
## The Challenge

Most medical image generation models require multiple high-end GPUs and massive datasets, putting them out of reach for smaller research labs. I wanted to see if I could build a functional text-to-X-ray generation system that could run on a single consumer GPU.

This was my exploration into making medical AI research more accessible - proving you don't need a supercomputer to experiment with generative models in healthcare.

## Technical Implementation

I built a latent diffusion model with three components:
- **VAE**: 26.2M parameters to compress X-rays into 8-channel latent space
- **UNet**: 39.66M parameters with cross-attention for denoising
- **BioBERT**: Fine-tuned text encoder (only 593K trainable parameters)

Total: 151.81M parameters, with just 43.5M trainable - about 10x smaller than comparable models.

The key was operating in latent space (32×32) instead of pixel space (256×256), reducing computation by 8x. I implemented gradient checkpointing, mixed precision training, and gradient accumulation to fit everything in 8GB VRAM.

## Training Process

Trained on the Indiana University Chest X-ray dataset (3,301 frontal images):
- **Stage 1**: VAE training for 67 epochs until convergence
- **Stage 2**: Diffusion model training for 387 epochs
- **Hardware**: Single RTX 4060 GPU (8GB VRAM)
- **Total time**: ~96 hours (vs 552+ hours for multi-GPU approaches)

The model achieves reasonable metrics (SSIM: 0.82, PSNR: 22.3dB) and generates 256×256 images in 663ms.

## What It Can (and Can't) Do

The model generates plausible-looking chest X-rays from text prompts like:
- "Normal chest X-ray with clear lungs"
- "Right lower lobe pneumonia with consolidation"
- "Cardiomegaly with pulmonary congestion"

But let's be clear about limitations:
- **No clinical validation** - I didn't have radiologists evaluate the outputs
- **Not for diagnostic use** - This is purely a research demonstration
- **Limited resolution** - 256×256 isn't sufficient for clinical detail
- **Single institution data** - Trained only on Indiana University dataset

## Real Impact

While this isn't clinically validated, it demonstrates that medical AI research doesn't require massive resources. The complete codebase, trained models, and detailed documentation are open source, allowing other researchers to:
- Experiment with medical image generation on consumer hardware
- Understand optimization techniques for resource-constrained training
- Build upon the architecture for their own research

## Lessons Learned

- Latent diffusion is incredibly efficient - 8x spatial compression makes consumer GPU training feasible
- Domain-specific design matters - 8 latent channels worked better than the standard 4 for X-rays
- Mixed precision + gradient checkpointing + accumulation can reduce memory by 65%
- You can do meaningful research without A100 GPUs

This project won't revolutionize radiology, but it might help democratize medical AI research by showing what's possible with limited resources.

*Paper preprint: https://www.preprints.org/manuscript/202506.1783/v1*  
*All code and models: https://github.com/priyam-choksi/cxr-diffusion*

Text-Conditional Chest X-ray Generation Using Latent Diffusion


![System Architecture](/images/project/citypipeline.jpg)

I built this pipeline to understand how cities handle massive amounts of sensor data in real-time. Think of it as the backend for those smart traffic systems that adjust lights based on traffic flow, or emergency systems that route ambulances through the fastest paths.

## What It Does

The pipeline ingests data from multiple sources:
- Vehicle data (speed, location, engine diagnostics)
- GPS tracking from city vehicles
- Weather sensor readings
- Traffic camera metadata
- Emergency system alerts

All this data flows through Kafka, gets processed by Spark, and lands in AWS for analysis. Pretty standard streaming architecture, but the interesting part was handling different data types and rates simultaneously.

## How I Built It

Started with Docker Compose to run everything locally. This was key - nobody wants to spin up a whole AWS cluster just to test something.

The main components:
- **Kafka cluster**: 3 brokers with Zookeeper for coordination
- **Spark streaming**: Processing engine that does the heavy lifting
- **Python scripts**: Generate realistic data and handle specific processing logic
- **AWS stack**: S3 for storage, Glue for cataloging, Athena for queries

### The Data Flow

1. Python producers generate data mimicking real sensors - vehicles moving through the city, weather changing, occasional emergency events
2. Kafka ingests everything, routing to different topics based on data type
3. Spark reads from Kafka, does windowed aggregations (5-minute traffic summaries, hourly weather patterns)
4. Processed data goes to S3, partitioned by date/hour for efficient querying
5. Glue crawlers keep the schema updated as new data arrives
6. Analysts can query everything through Athena using regular SQL

## Interesting Challenges

**Different data rates**: Weather updates every minute, vehicles send data every second, cameras stream continuously. Had to tune Kafka partitions and consumer groups to handle this without one slow source blocking others.

**State management**: Tracking vehicles across zones required maintaining state in Spark. Used mapWithState for efficiency - way better than updateStateByKey.

**Schema evolution**: Sensors get upgraded, new fields appear. Built the pipeline to handle schema changes gracefully instead of breaking.

**Cost optimization**: S3 gets expensive fast. Implemented smart partitioning and lifecycle rules to move old data to cheaper storage.

## What I Learned

- Kafka's partition strategy matters more than you think
- Spark Streaming's micro-batches are actually pretty clever for this use case
- Docker Compose is amazing for local development of distributed systems
- AWS Glue is convenient but can get pricey if you're not careful with crawler schedules

## Real-World Use

This architecture works for any multi-source streaming scenario:
- Fleet management systems
- Industrial IoT monitoring  
- Real-time analytics platforms
- Event-driven microservices

The patterns are the same whether you're processing city sensors or e-commerce clickstreams.

*Code and setup instructions: https://github.com/priyam-choksi/City-Real-Time-Streaming-Data-Pipeline*

Smart City Real-Time Data Streaming Pipeline


![System Architecture](/images/project/dwbi.jpeg)

A full data warehouse implementation using the AdventureWorks dataset, built to simulate the kind of multi-source, multi-region data environments common in enterprise settings. The goal was to consolidate data from four different database platforms into a single analytical layer that supports meaningful business reporting.

## Architecture

Data is sourced from four regional systems running on different databases:

- SQL Server (US region)
- MySQL (EU region)
- PostgreSQL (AU region)
- Oracle (CA region)

All four feed into a central data warehouse — `AdventureWorksDW_neu` — built on both MySQL (Talend target) and SQL Server (SSIS target).

## ETL Pipeline

Built parallel ETL pipelines in Talend and SSIS to cover both toolsets:

- **Extraction**: Connected to all four source systems with error handling at each step
- **Transformation**: Applied business rules, data cleansing, and standardization across regions
- **Loading**: Optimized bulk loading with indexing strategies tuned for analytical query patterns

## Data Model

Designed a star schema to support sales and purchasing analysis:

- **Dimension tables**: DimStore, DimCustomerPerson, DimCurrency, DimPromotion, DimSalesTerritory, and others
- **Fact tables**: FactInternetSales, FactStoreSales, FactPurchases
- **Slowly Changing Dimensions**: Type 1 implementation for handling business changes over time
- **Reject tables**: Capturing records that fail validation, with reconciliation reports

## Business Intelligence Layer

Built a reporting layer on top of the warehouse:

- Materialized views for frequently queried patterns
- Vendor ranking by purchase volume
- Sales performance by territory, product, and channel
- Employee and vendor involvement in purchasing workflows
- Dashboard-ready output structured for Power BI consumption

## What This Covers

- Multi-platform data integration across heterogeneous source systems
- ETL development in both Talend and SSIS
- Dimensional modeling with proper fact and dimension separation
- Data quality management with validation, rejection, and reconciliation
- BI query design for operational and analytical reporting

Data Warehouse Integration and Business Intelligence


During my internship at Brigham and Women's Hospital, I worked with the pharmacology team to fix their painfully slow genomic analysis pipeline. Researchers were waiting 2-3 days for results, which was killing their productivity and frustrating everyone involved.

## The Problem

The team was analyzing 5TB+ of multi-omics data - RNA sequencing, protein expression, metabolomics, plus clinical metadata. Their pipeline was a collection of bash scripts that:
- Processed everything sequentially (one sample at a time)
- Used text files for massive joins
- Requested way more cluster resources than needed (500GB memory for 50GB jobs)
- Could only run on the ERIS cluster, no local testing possible

Between the inefficient code and cluster queue times, researchers waited 48-72 hours just to find out if their analysis worked. If it failed, that's another 2-3 days down the drain.

## What I Did

First, I profiled the existing pipeline to understand where time was being wasted. Turns out, most of the "computation time" was actually:
- Waiting in queue due to overestimated resource requests
- Reading/writing massive text files repeatedly
- Processing samples one by one instead of in parallel

I rebuilt the pipeline using PySpark and switched the storage format to Parquet:
- **Parquet files**: 60% smaller than text files, way faster for column-based queries
- **Parallel processing**: Partitioned by sample ID so multiple samples process simultaneously
- **Right-sized resources**: Profiled actual memory usage and requested appropriate amounts
- **Local testing**: Set up Jupyter notebooks so researchers could test locally before cluster submission

## The Results

The improvements were dramatic:
- Processing time: 48-72 hours → 15 minutes
- Queue wait time: 2-6 hours → 5-10 minutes  
- Storage needs: Reduced by 60%
- Researcher productivity: From 2-3 analyses per week to 10-15

The 15-minute runtime includes all the joins, transformations, and statistical calculations. The speedup came from:
- Parallel processing (biggest win)
- Columnar storage format (Parquet)
- Eliminating redundant I/O operations
- Actually using the cluster's capabilities

## What Made This Work

The key insight was that this wasn't really a "big data" problem - it was a "badly organized data" problem. The cluster had plenty of power; the pipeline just wasn't using it properly.

I also learned that scientists don't care about fancy features - they want reliability and speed. So I kept the interface simple, wrote clear documentation, and made sure they could modify the pipeline themselves after I left.

## Technical Details

- **Data**: 1.3TB of genomic and clinical data
- **Cluster**: Harvard Medical School's ERIS HPC cluster
- **Old stack**: Bash scripts + text files + sequential processing
- **New stack**: PySpark + Parquet + parallel processing + Jupyter
- **Optimization**: Proper resource allocation, partition strategy, columnar storage

By the end of my internship, the team could iterate on analyses in real-time instead of waiting days. They were exploring hypotheses they'd previously avoided because the computational cost was too high.

*Note: Specific code and implementation details remain confidential per BWH/Harvard Medical School policies.*

Multi-omics Data Processing Pipeline


![System Architecture](/images/project/scalablepipeline.png)

I built this pipeline to get hands-on experience with the tools that power data infrastructure at companies like Netflix and Uber. It's a complete implementation showing how different technologies work together to create a resilient streaming data system.

## What I Built

A fully functional data pipeline that:
- Ingests data through Airflow-orchestrated jobs
- Streams it through Kafka for reliable delivery
- Processes with Spark for transformations
- Stores in Cassandra for high-speed access

The interesting part wasn't just getting these tools to work individually - it was making them work together seamlessly. Each component has its quirks and getting them to play nice required understanding their internals.

## Technical Implementation

### The Stack

**Apache Airflow**: Orchestrates the whole pipeline. I wrote DAGs that:
- Fetch data from APIs (using randomuser.me for demo data)
- Handle retries and failures gracefully
- Monitor pipeline health

**Kafka + Zookeeper**: The nervous system of the pipeline
- Configured multi-broker setup for fault tolerance
- Tuned partition strategy for parallel processing
- Implemented proper serialization for data integrity

**Spark Streaming**: Where the magic happens
- Built streaming jobs that process data in micro-batches
- Implemented transformations and aggregations
- Set up checkpointing for failure recovery

**Cassandra**: The final destination
- Designed tables optimized for write throughput
- Denormalized data model for fast reads
- Configured replication for high availability

### Docker Everything

The entire stack runs in Docker containers. This was crucial because:
- Anyone can spin up the whole pipeline with one command
- No "works on my machine" problems
- Easy to experiment without breaking things

Getting the networking right was tricky - Kafka needs to advertise correct addresses, Spark workers need to find the master, and everything needs to reach Cassandra.

## Challenges and Solutions

**Kafka's Zookeeper dependency**: Understanding why Kafka needs Zookeeper taught me about distributed consensus. When Zookeeper went down during testing, Kafka became headless - a good lesson in failure modes.

**Spark memory management**: Initial runs kept failing with OOM errors. Learned to tune executor memory, understand the difference between driver and executor, and how to partition data properly.

**Cassandra consistency levels**: Choosing between consistency and availability isn't theoretical when your writes start failing. Settled on LOCAL_QUORUM for the right balance.

## Performance Insights

The pipeline can handle significant load on modest hardware:
- Processes thousands of events per second
- Scales horizontally by adding Kafka partitions and Spark workers
- Cassandra handles writes without breaking a sweat

But more importantly, I learned where the bottlenecks appear:
- Network I/O between Docker containers
- Serialization/deserialization overhead
- Checkpoint storage for Spark streaming

## Real Learning

This project taught me that distributed systems are hard not because of any single component, but because of the interactions between them. Some key lessons:

- **Start simple**: My first version just moved data from A to B. Added complexity gradually.
- **Monitor everything**: Added Prometheus + Grafana to understand what's actually happening
- **Failure is normal**: Every component will fail. Design for it.
- **Documentation matters**: Future me thanked past me for writing down configuration decisions

## Practical Applications

While this uses synthetic data, the architecture applies to:
- E-commerce event streams
- IoT sensor networks
- Financial transaction processing
- Log aggregation systems

The patterns are the same whether you're processing user clicks or temperature readings.

*Full implementation with setup instructions and monitoring dashboards: [GitHub](https://github.com/priyam-choksi/Scalable-Data-Engineering-Pipeline-using-Apache-Kafka-Apache-Spark-and-Cassandra)*

Scalable Data Pipeline with Apache Airflow, Kafka, and Cassandra


![System Architecture](/images/project/stockpipeline.jpg)

Designed and implemented a real-time data engineering pipeline that demonstrates the same architectural patterns used by financial institutions for processing market data streams. The system handles continuous data ingestion, processing, and analytics-ready storage using industry-standard technologies.

## Architecture Overview

Built a multi-layer streaming architecture that separates concerns across:
- **Ingestion Layer**: Kafka producers simulating real-time market feeds
- **Streaming Layer**: Kafka brokers on EC2 handling message distribution
- **Processing Layer**: Python consumers with configurable batch processing
- **Storage Layer**: S3 with partitioned data organization for query optimization
- **Analytics Layer**: Glue + Athena enabling SQL analytics on streaming data

This architecture mirrors production systems while remaining cost-effective for demonstration purposes.

## Technical Implementation

### Kafka Configuration
Deployed single-node Kafka on EC2 with custom configurations:
- Topic partitioning strategy for parallel processing
- Retention policies balancing storage and data availability
- Producer configs optimizing for throughput vs latency

### Data Pipeline Components

**Producer Application** (`KafkaProducer.ipynb`)
- Implements configurable data generation rates
- Handles connection failures with exponential backoff
- Serializes market data maintaining schema consistency

**Consumer Application** (`KafkaConsumer.ipynb`)
- Batch processing for S3 write optimization
- Implements at-least-once delivery semantics
- Transforms data into queryable Parquet format

### AWS Integration

Leveraged AWS services for scalable data infrastructure:
- **S3**: Organized with date-based partitioning for efficient queries
- **Glue Crawler**: Automated schema discovery and catalog updates
- **Athena**: Enabled ad-hoc SQL analysis on streaming data

## Engineering Decisions

**Why Kafka over AWS Kinesis**: Chose Kafka for deeper understanding of streaming internals and portability across cloud providers.

**S3 Partitioning Strategy**: Implemented YYYY/MM/DD/HH partitioning to optimize for time-based queries typical in financial analysis.

**Data Format**: Selected Parquet for columnar storage benefits - 70% storage reduction compared to JSON while improving query performance.

## Performance Characteristics

The pipeline successfully demonstrates:
- Sub-second latency from producer to Kafka
- Configurable batch intervals (1-60 seconds) for S3 writes
- Athena queries returning in seconds on gigabytes of data
- Horizontal scalability through Kafka partitions

## Practical Applications

This architecture pattern applies to:
- Real-time trading systems
- Risk management platforms
- Market surveillance systems
- Regulatory reporting pipelines

The same principles scale to handle millions of messages per second with appropriate infrastructure.

## Key Takeaways

Building this pipeline provided hands-on experience with:
- Distributed systems challenges (partition management, consumer group coordination)
- Stream processing patterns (windowing, watermarking concepts)
- Cost-performance tradeoffs in cloud architectures
- Data lake best practices for analytics workloads

While implemented as a single-node demo, the architecture supports horizontal scaling by adding Kafka brokers and consumer instances.

*Complete implementation with setup documentation: https://github.com/priyam-choksi/Real-Time-Stock-Market-Data-Processing*

Real-Time Stock Market Data Processing with Kafka and AWS


![System Architecture](/images/project/uberarch.jpg)

I built this pipeline to explore Google Cloud Platform's data ecosystem while analyzing NYC taxi data. Using trip records from yellow and green cabs, the project demonstrates how modern cloud tools can transform raw transportation data into actionable insights about urban mobility patterns.

## The Dataset and Challenge

The NYC Taxi and Limousine Commission publishes detailed trip data - pickup/dropoff locations, times, distances, fares, and payment types. It's perfect for understanding urban transportation patterns, but the volume makes traditional analysis tools struggle.

The challenge was building a pipeline that could process this data efficiently while staying maintainable and cost-effective. I wanted to answer questions like: How do trip patterns change during rush hours? Which zones generate the most revenue? How do weather events impact taxi demand?

## Building with Modern Tools

I chose Mage as the orchestration tool after evaluating Airflow and Prefect. Mage's visual pipeline builder and built-in data quality checks made development significantly faster. Plus, I wanted to contribute back to their open-source project, which I did by improving their BigQuery connector documentation.

The architecture follows a classic medallion pattern. Raw CSV files land in Google Cloud Storage, Mage orchestrates the transformation pipeline running on Compute Engine, cleaned data flows into BigQuery, and Looker Studio provides the visualization layer. Each component was chosen for specific strengths - GCS for cheap blob storage, BigQuery for scalable analytics, Looker for self-service dashboards.

## Technical Implementation

The ETL pipeline handles several complex transformations. Parsing NYC's location IDs required joining with their official zone lookup tables. Calculating trip durations revealed data quality issues - some trips showed negative times due to system errors. I implemented data quality checks that flag anomalies while preserving valid edge cases.

BigQuery's partitioning by date and clustering by pickup location reduced query costs significantly. A query scanning the full dataset would cost dollars; with proper partitioning, it's pennies. This optimization was crucial for making the dashboards responsive and affordable.

The Looker Studio dashboards update automatically as new data arrives. Users can filter by date ranges, zones, or payment types to understand trends. One interesting insight: credit card payments have steadily increased, but cash still dominates in certain neighborhoods, revealing digital divide patterns across the city.

## Lessons from Cloud Migration

This project taught me practical lessons about cloud architectures. BigQuery's serverless model removes infrastructure headaches but requires thinking differently about cost optimization. Instead of provisioning servers, you optimize queries and data organization.

I also learned the importance of incremental processing. Initial runs processed all historical data, but daily updates only process new records. This pattern, implemented through Mage's incremental models, keeps costs linear as data grows.

## Real-World Applications

While this uses taxi data, the patterns apply broadly. The same architecture could analyze e-commerce transactions, IoT sensor readings, or application logs. The key is choosing the right tool for each job - orchestration for workflows, columnar stores for analytics, visualization tools for insights.

Contributing to Mage's open-source project was rewarding. My documentation improvements help others avoid the BigQuery authentication issues I struggled with. Open source isn't just about taking; it's about improving tools for everyone.

This pipeline serves as a reference implementation for organizations migrating analytics workloads to GCP. It demonstrates that cloud-native doesn't mean complex - sometimes the simplest architecture is the most powerful.

*Full implementation and cost analysis available on GitHub.*

Uber Data Analytics Pipeline on Google Cloud Platform


![Web App Screenshot](/images/project/diabetes.png)

Built a full-featured diabetes prediction platform that combines machine learning with interactive data analysis. The project started as part of my capstone but evolved into a comprehensive tool with separate modules for prediction, exploratory data analysis, and educational visualizations.

## Technical Overview

The platform processes clinical data including glucose levels, BMI, HbA1c, age, and lifestyle factors to predict diabetes risk. I evaluated five different ML algorithms before selecting logistic regression for its interpretability in medical contexts.

### Key Components:
- **Prediction Engine**: Logistic regression model achieving 95.97% accuracy and 0.9587 ROC-AUC
- **Interactive Dashboard**: Real-time predictions with visual feedback on health metrics
- **EDA Module**: Statistical analysis and visualizations of the 100,000-record training dataset
- **Educational Features**: Shows how each health metric influences the prediction

## Implementation Details

The interesting challenges came from balancing model performance with interpretability. While gradient boosting models achieved slightly higher accuracy, logistic regression's coefficient interpretation was crucial for helping users understand their risk factors.

I implemented comprehensive data preprocessing pipelines using StandardScaler for numerical features and OneHotEncoder for categorical variables. The model uses L2 regularization to prevent overfitting on the large dataset.

The Streamlit interface required careful optimization - caching the model load, managing session states for multiple users, and ensuring responsive performance even with complex visualizations running simultaneously.

## Research and Validation

This wasn't just a coding project - I wrote a full research paper comparing different algorithms:
- **Logistic Regression**: Selected for interpretability and consistent performance
- **SVM**: Similar accuracy but black-box nature made it unsuitable
- **Random Forest**: Marginally better performance but harder to explain to users
- **XGBoost**: Best raw accuracy but computational overhead wasn't justified
- **KNN**: Poor performance on this high-dimensional medical data

The mathematical foundations were important - understanding why the sigmoid function works for probability estimation and how regularization prevents overfitting on medical data with many correlated features.

## Lessons Learned

Medical ML is different from typical ML projects. Interpretability matters more than raw accuracy. A 96% accurate model that doctors can understand beats a 98% black box. The project taught me about the regulatory and ethical considerations in healthcare technology.

The full implementation includes proper error handling, input validation for medical ranges, and clear disclaimers about not replacing professional medical advice.

## Links

- **Live App**: https://diabetes-prediction-and-analytics.streamlit.app/
- **GitHub**: https://github.com/priyam-choksi/Diabetes-Streamlit-App/
- **Research Paper**: Available in the repository
- **Demo Video**: https://youtu.be/gIcCXWj7e-A

Diabetes Prediction and Analysis Platform


## Project Overview

The "DS-ML Notebooks" repository is a collection of Jupyter Notebooks focused on Data Science and Machine Learning topics. This repository serves as a central hub for various notebooks created for Kaggle competitions, personal projects, and other learning exercises. Each notebook is designed to be a hands-on example of applying data science and machine learning techniques to real-world datasets.

### **Key Features**

- **Comprehensive Coverage:** The repository includes notebooks covering a wide range of topics in data science and machine learning, from basic data manipulation to advanced deep learning techniques.
- **Kaggle Notebooks:** A significant portion of the notebooks are dedicated to Kaggle competitions, providing practical examples of how to approach and solve data science challenges.
- **Educational Value:** Each notebook is crafted with educational purposes in mind, making it a valuable resource for learners and practitioners looking to enhance their skills.
- **Reproducibility:** All notebooks are designed to be easily reproducible, with clear explanations, code snippets, and well-commented sections.
- **Real-World Application:** The notebooks apply theoretical concepts to real-world datasets, showcasing the practical utility of data science and machine learning techniques.

### **Technologies Used**

- **Python:** The primary programming language used across all notebooks.
- **Jupyter Notebooks:** The interactive environment for creating and sharing code, equations, visualizations, and narrative text.
- **Pandas:** For data manipulation and analysis, providing tools to clean, transform, and summarize data.
- **NumPy:** For numerical computations and array processing, essential for scientific computing in Python.
- **Scikit-learn:** A machine learning library used for implementing various algorithms, including regression, classification, clustering, and model evaluation techniques.
- **TensorFlow & Keras:** Deep learning frameworks used for building and training neural networks, focusing on tasks such as image recognition, natural language processing, and more.
- **Matplotlib & Seaborn:** Libraries for creating static, animated, and interactive visualizations in Python, helping to uncover patterns and insights within the data.

### **Repository Highlights**

- **Exploratory Data Analysis (EDA):** Notebooks that delve into exploratory data analysis, showcasing techniques to understand the underlying structure of the data, detect anomalies, and identify significant variables.
- **Feature Engineering:** Practical examples of feature engineering, demonstrating how to transform raw data into features that better represent the underlying problem to predictive models.
- **Modeling:** Includes notebooks on building and tuning machine learning models, from simple linear regressions to complex deep learning models, with a focus on improving model accuracy and robustness.
- **Hyperparameter Tuning:** Demonstrates methods for optimizing model performance through techniques such as grid search, random search, and Bayesian optimization.
- **Visualization Techniques:** Notebooks dedicated to advanced data visualization techniques, including interactive plots, geospatial visualizations, and time-series analysis.
- **Experimentation and Validation:** Focused on model validation techniques, such as cross-validation, hold-out validation, and A/B testing, to ensure that the models generalize well to unseen data.

DS-ML Notebooks


![Web App Screenshot](/images/project/Ecotrip.png)

#### Overview
**EcoTrip** is a mobile application designed to encourage and facilitate sustainable travel. The app targets eco-conscious travelers, providing them with tools to make environmentally responsible travel choices, book eco-friendly accommodations, and engage with local sustainable businesses.

#### Objective
The primary objective of EcoTrip was to fill a gap in the travel industry by providing a platform that not only assists in travel planning but also promotes sustainability through its recommendations and features.

#### My Role
My responsibilities included user research, designing the user interface and experience, prototyping, and overseeing the user testing phases to refine the application.

---

#### Process

1. **Research**
   - **Market Analysis**: Conducted an extensive review of existing travel apps to identify missing features and areas for improvement, particularly in promoting sustainable practices.
   - **Target Audience**: Identified our primary users as environmentally conscious millennials who prefer making travel choices that align with their values.

2. **Design**
   - **User Personas**: Created detailed personas such as "Eco-friendly Emma", a 29-year-old who looks for the most sustainable travel options.
   - **User Journey Maps**: Developed journey maps to understand potential frustrations and needs during the travel planning process.
   - **Wireframing**: Sketched out initial wireframes on paper and then digitized them using Adobe XD.

3. **Prototyping and Testing**
   - **Interactive Prototyping**: Moved from static wireframes to interactive prototypes in Figma, which allowed for the simulation of user interactions with the interface.
   - **User Testing**: Conducted multiple rounds of user testing with prototypes to gather feedback and refine the UI/UX based on real user data.

---

#### Tools & Technologies
- **Figma**: Used for high-fidelity prototyping and interactive features.
- **Adobe XD**: Employed for initial wireframing and mockups.
- **UserTesting**: Utilized to conduct remote user testing sessions and gather feedback.

---

#### Design Highlights
- **Sustainability Calculator**: Integrated a feature that calculates the carbon footprint of various travel options, helping users make informed decisions.
- **Green Badge System**: Designed a reward system where users earn badges for making eco-friendly choices, enhancing user engagement.
- **Accessible and Inclusive Design**: Ensured the app's design met accessibility standards to accommodate all users, including those with disabilities.

---

#### Challenges and Solutions
- **Challenge**: Balancing aesthetic appeal with the informative content necessary for promoting sustainability.
- **Solution**: Adopted a clean and minimalistic design approach, using visual elements that resonate with eco-friendly aesthetics (e.g., earth tones and natural imagery).

---

#### Outcomes and Impact
- **User Engagement**: Post-launch metrics indicated a 40% increase in app registrations, with users citing the eco-features as a major draw.
- **Business Impact**: Partnered with eco-friendly businesses which reported increased visibility and customer engagement through the app.
- **Awards and Recognition**: EcoTrip received a notable design award for innovation in sustainability at a well-known tech conference.

---

#### Lessons Learned
- The importance of iterative design and user feedback was reinforced, highlighting how vital these elements are in creating a product that truly resonates with users.
- Gained deeper insights into designing for sustainability, which requires a careful balance between user expectations and environmental advocacy.

---

#### Conclusion
**EcoTrip** has set a new benchmark for sustainable travel apps, proving that thoughtful design can lead to meaningful changes in consumer behavior. This project not only enhanced my skills in user-centered design but also deepened my commitment to sustainability in tech.

EcoTrip: Sustainable Travel App Case Study


**Introduction:**

In this innovative project, I applied advanced machine learning techniques to predict outcomes in Formula 1 races, aiming to provide teams with strategic insights that can enhance their competitive edge. By analyzing historical race data, driver performance, and environmental conditions, I developed a predictive model that offers high accuracy in forecasting race results.

**What I Did:**

- **Data Collection and Preprocessing:** Gathered a comprehensive dataset including drivers' historical performances, track characteristics, weather conditions, and car specifications. This data was meticulously cleaned and prepared for analysis, ensuring a robust foundation for the models.

- **Feature Engineering:** Engineered predictive features that significantly impact race outcomes, such as tire wear, pit stop strategy, and qualifying positions. This step involved extensive analysis and transformation of raw data into actionable insights.

- **Model Development:** Utilized TensorFlow and Scikit-learn to build and train several predictive models, including decision trees, random forests, and neural networks. These models were chosen for their ability to handle complex, non-linear relationships within the data.

- **Model Optimization and Validation:** Conducted rigorous testing and validation procedures to fine-tune the models. This process included cross-validation and hyperparameter tuning to optimize performance and prevent overfitting.

- **Deployment and Real-time Testing:** Integrated the model into a real-time analytics tool used by an F1 team to make informed decisions during race weekends. This tool provides predictions on race strategies and potential outcomes, aiding in crucial decision-making processes.

**What I Learned:**

- **Deep Dive into Sports Analytics:** Gained deep insights into the specific challenges and requirements of sports analytics, particularly in the fast-paced environment of Formula 1 racing.

- **Advanced Machine Learning Applications:** Enhanced my expertise in applying complex machine learning algorithms to real-world problems, learning to adapt and tailor models according to specific domain needs.

- **Impact of Analytics in Competitive Sports:** Learned about the transformative impact that data analytics and machine learning can have on sports, driving strategies that can alter the outcomes of high-stakes competitions.

**Key Insights:**

- **Strategic Advantage:** Demonstrated how predictive analytics can provide a significant strategic advantage in Formula 1 by optimizing race strategies and predicting competitor moves.

- **Enhanced Decision-Making:** Highlighted the role of data-driven decisions in improving race day strategies, which can lead to better placements and potentially winning more races.

- **Future of Sports Analytics:** Showed the potential for further integration of machine learning in sports, suggesting that data-driven approaches could become central to developing future sporting strategies and technologies.

**Summary:**

The "Predicting F1 Racing Outcomes with Machine Learning" project showcases the powerful combination of sports and technology, where advanced data analytics significantly influence competitive strategies. This initiative has not only propelled my growth as a data scientist but also contributed to the evolution of strategy development in Formula 1 racing, paving the way for more innovative uses of technology in sports analytics.

---

Predicting F1 Racing Outcomes with Machine Learning

![Web App Screenshot](/images/project/child.png)

[View Figma Exports (PDF)](/childadoption.pdf){:target="_blank"}


## Introduction:

“Forever Home” is an innovative digital platform designed to simplify the process of child adoption. This project focuses on creating a user-friendly interface that connects prospective adoptive parents with children in need of loving homes. By integrating essential adoption services into one comprehensive application, “Forever Home” aims to streamline the adoption process while ensuring the best possible outcomes for both children and families.

## What I Did:

- **User-Centered Design Process:** Implemented a comprehensive design thinking approach, including empathizing with users, defining the problem, ideating solutions, prototyping, and testing. This process ensured that the final product meets the specific needs of both adoptive parents and children.
  
- **Extensive User Research:** Conducted detailed user research to understand the needs, preferences, and challenges of both adoptive parents and children. This research informed the design process and guided the development of key features.
  
- **SWOT Analysis:** Performed a SWOT analysis to identify the strengths, weaknesses, opportunities, and threats in the adoption process. This analysis helped in shaping strategic decisions in the app’s development and feature prioritization.
  
- **High-Fidelity Prototyping:** Utilized Figma and Adobe XD for high-fidelity prototyping, creating interactive prototypes that showcase the app's user interface, navigation flows, and key features.

## What I Learned:

- **Advanced Prototyping and Wireframing Skills:** Enhanced my proficiency in using Figma and Adobe XD to create dynamic prototypes and wireframes that effectively communicate design concepts and facilitate user testing.
  
- **Strategic Design Implementation:** Gained insights into strategic planning in design, learning how to align user needs with business objectives using methods like the MoSCoW approach to prioritize features throughout the design and development phases.
  
- **Empathetic User Engagement:** Improved my ability to engage with users empathetically, an essential skill in designing solutions that address real-world problems in child adoption, enhancing user satisfaction and overall usability of the platform.

## Key Insights:

- **Simplifying the Adoption Process:** The project highlighted the importance of simplifying the adoption process through intuitive design and user-friendly features, making it easier for prospective parents to navigate the complexities of adoption.
  
- **Impact of User Filters and Search:** Demonstrated how the inclusion of user filters and search options can significantly improve the adoption process, allowing users to find children that meet their specific criteria and preferences.

## Summary:

The “Forever Home” project exemplifies the power of thoughtful UI/UX design in transforming child adoption solutions. Through meticulous user research, empathetic design practices, and strategic feature implementation, the project not only simplifies the adoption process but also creates a platform that promotes the well-being and happiness of children in need of loving homes. This initiative has set a new standard in the adoption industry, proving that design can play a crucial role in creating positive social impact.


Forever Home: A Child Adoption App Transforming Lives

**Introduction:**

In an innovative blend of art and technology, this project aimed to revolutionize the way we interact with digital media by developing a system capable of applying the style of one image to the content of another. Using Generative Adversarial Networks (GANs), this initiative pushed the boundaries of conventional image processing to create novel and visually striking artistic images.

**What I Did:**

- **Model Design and Implementation:** Led the development and training of GANs using TensorFlow and Keras. This involved designing the architecture to effectively capture and replicate diverse artistic styles, from classic to contemporary.

- **Hyperparameter Optimization:** Undertook meticulous tuning of the model parameters to strike an optimal balance between style fidelity and image clarity. This was crucial in ensuring that the style transfers were both visually appealing and recognizably accurate.

- **Interactive Web Interface Development:** Engineered a user-friendly web interface that allows users to seamlessly upload their images and experiment with different artistic styles in real-time. This interface was designed to be intuitive, catering to both enthusiasts and professional designers.

**What I Learned:**

- **Advanced Deep Learning Techniques:** Gained profound insights into the workings of deep learning frameworks, particularly GANs, enhancing my understanding of neural network training and stabilization techniques.

- **User-Centric Design:** The project honed my skills in developing user-centered applications, emphasizing interactive design and real-time feedback mechanisms.

- **End-to-End Project Management:** From conceptualization to deployment, I managed all aspects of this project, which improved my abilities in project planning, execution, and cross-functional collaboration.

**Key Insights:**

- **Innovative Artistic Creation:** Demonstrated the potential of GANs to transform digital art, enabling the creation of unique images that blend multiple artistic influences.

- **Enhanced Visual Communication:** The ability to manipulate image styles on-the-fly can serve as a powerful tool for designers, advertisers, and content creators, offering new ways to engage and captivate audiences.

- **Technological Integration:** Highlighted the importance of integrating advanced machine learning techniques with user-friendly interfaces to make cutting-edge technology accessible to a broader audience.

**Summary:**

The "Image Style Transfer Using GANs" project represents a significant leap forward in digital media manipulation, combining deep learning technology with artistic creativity to produce a versatile tool for image styling. This venture not only enhanced my technical proficiency in handling complex neural networks but also reinforced my commitment to integrating innovative technology with practical applications, ensuring that advancements in AI continue to enrich our visual and interactive experiences.


Image Style Transfer Using GANs


# HealthHub: Personal Health Management App

## Overview
HealthHub is a comprehensive personal health management app designed to empower users to take control of their health by tracking, analyzing, and managing their medical and wellness data. It aims to provide a centralized platform where users can easily monitor various aspects of their health, from daily activity levels to detailed medical records, and receive personalized insights and recommendations.

## Objective
The primary objective of HealthHub is to create a user-centric platform that enables individuals to efficiently manage their health-related data and receive personalized recommendations, thereby fostering proactive health management and improving overall well-being.

## My Role
As the lead UI/UX designer, my responsibilities encompassed the entire design process, from conducting comprehensive user research to refining the application through iterative testing. This included developing user personas, crafting intuitive user flows, designing accessible interfaces, and creating interactive prototypes for user testing.

## Process
### Research
- **Stakeholder Interviews:**
  Conducted in-depth interviews with healthcare professionals and potential users to gain insights into their needs, preferences, and pain points regarding health management applications.
- **Competitive Analysis:**
  Analyzed existing healthcare apps to identify common features, assess user experience, and pinpoint areas for innovation and improvement within the HealthHub ecosystem.
  
### Design
- **User Personas:**
  Developed detailed user personas, such as "Diabetic Dave," to represent diverse user demographics and tailor design decisions to meet specific user needs, such as medication management and dietary tracking.
- **User Flow and Wireframes:**
  Created comprehensive user flows and wireframes to visualize the app's navigation structure and screen layouts, ensuring a seamless user experience and efficient task completion.
- **Accessibility Considerations:**
  Prioritized accessibility by adhering to ADA standards, including considerations for readability, navigation simplicity, and interaction clarity to accommodate users with disabilities and diverse needs.
  
### Prototyping and Testing
- **High-Fidelity Prototyping:**
  Utilized Sketch and InVision to develop high-fidelity prototypes that closely resembled the final application, enabling stakeholders to visualize the user interface and interactions in detail.
- **User Testing:**
  Organized multiple rounds of user testing sessions to gather feedback on prototype iterations, validate design decisions, and refine the UI/UX based on user insights, ensuring usability and satisfaction.

## Tools & Technologies
- Sketch
- InVision
- SurveyMonkey

## Design Highlights
- **Health Dashboard:**
  Designed an intuitive dashboard that aggregates key health metrics, such as blood pressure, glucose levels, and calorie intake, providing users with a comprehensive overview of their health status at a glance.
- **Medication Tracker:**
  Integrated a robust medication management feature, including medication schedules, reminders, and adherence tracking, to support users in managing their prescriptions effectively and promoting medication compliance.
- **Personalized Health Insights:**
  Implemented an AI-driven insights engine that analyzes user data, including activity levels, medical history, and lifestyle habits, to generate personalized health recommendations and alerts tailored to individual user profiles.

## Challenges and Solutions
- **Challenge: Integrating complex medical data in a user-friendly manner.**
  - **Solution:** Developed a modular design approach that allows users to customize their dashboard layout and prioritize health metrics relevant to their specific conditions, enhancing usability and information accessibility.

## Outcomes and Impact
- **User Adoption:**
  Achieved a significant 50% increase in monthly active users within six months post-launch, indicating a strong user engagement and satisfaction with the HealthHub platform.
- **Clinical Partnerships:**
  Established strategic partnerships with several clinical institutions to enrich the app with expert insights, evidence-based content, and professional guidance, enhancing the app's credibility and value proposition for users.

## Lessons Learned
- **Importance of Data Security and Privacy:**
  Recognized the critical significance of prioritizing data security and privacy measures in health-related applications to ensure user trust, compliance with regulations, and safeguarding sensitive personal health information.
- **Flexibility in Design Frameworks:**
  Acquired valuable skills in designing flexible, user-centered design frameworks capable of accommodating diverse user needs, preferences, and health conditions, fostering inclusivity and usability across varied user demographics.


HealthHub: Personal Health Management App


![Integrated Healthcare System Screenshot](/images/project/ihcs.png)

This project represents the culmination of efforts to design a sophisticated healthcare management platform. The Integrated Healthcare System consolidates various healthcare components—hospitals, departments, doctors, patients, and more—into a unified Java-based system. This project's aim is to enhance communication and coordination across the healthcare sector, thereby improving operational efficiency and patient care.

**What did I do?**  
I led the development of this comprehensive platform, focusing on the integration of disparate healthcare services. My responsibilities included designing the system architecture, coding in Java, and ensuring seamless data integration using the db4o database system. I also implemented the Faker module for generating synthetic data, crucial for testing and validating system functionalities.

**What I used?**

- **[Java](https://www.oracle.com/java/):** The core programming language for developing the Integrated Healthcare System, chosen for its robustness, object-oriented features, and wide adoption in enterprise environments.

- **[Faker](https://github.com/DiUS/java-faker):** A Java library used to generate fake data for testing. This tool was instrumental in simulating realistic healthcare scenarios to test system performance and reliability.

- **[db4o](http://www.db4o.com/):** An object-oriented database used for managing complex data relationships within the system. Its integration allows for efficient data storage, retrieval, and manipulation, supporting the complex data structures of the healthcare system.

- **[Eclipse](https://www.eclipse.org/)** and **[IntelliJ IDEA](https://www.jetbrains.com/idea/):** Integrated Development Environments (IDEs) that provided comprehensive facilities for software development. I used these tools for coding, debugging, and managing the project effectively.

- [Machine Learning](https://en.wikipedia.org/wiki/Machine_learning) - Employed machine learning algorithms to enhance decision-making processes and automate tasks within the Integrated Healthcare System. These algorithms enabled predictive analytics and personalized patient care, contributing to improved efficiency and effectiveness.


**What I learned?**  
Through this project, I deepened my understanding of Java application development in a real-world context, enhanced my skills in database management, and learned to handle the complexities of an integrated system. The experience sharpened my problem-solving skills, particularly in the areas of system design and data integration. Additionally, I gained valuable insights into the healthcare industry's operational needs, which will be beneficial for future projects aimed at transforming other sectors.


Integrated Healthcare System

**Introduction:**

This project taps into the power of Long Short-Term Memory (LSTM) networks to forecast weather conditions with a high degree of accuracy. Utilizing historical weather data, this deep learning approach improves the predictive capabilities of traditional forecasting models by effectively capturing and analyzing complex patterns in meteorological time-series data.

**What I Did:**

- **Data Preparation:** Collected extensive historical weather datasets from meteorological stations, focusing on variables like temperature, humidity, atmospheric pressure, and wind data. This data underwent rigorous preprocessing to ensure it was clean and structured for effective model training.

- **LSTM Model Development:** Developed LSTM models using TensorFlow and Keras to predict future weather conditions. These models were specifically chosen for their ability to remember information for long periods, which is essential for accurate weather prediction.

- **Model Training and Testing:** Trained the models on historical data and conducted thorough testing to assess their accuracy and reliability. The models were fine-tuned through various iterations to enhance their forecasting performance.

- **Evaluation:** Evaluated the models against existing benchmarks and statistical methods to validate their superiority in handling sequential data and predicting dynamic weather patterns.

**What I Learned:**

- **Advanced LSTM Applications:** Enhanced my understanding of LSTM networks and their application in handling sequential data, particularly in the challenging domain of weather forecasting.

- **Model Optimization Techniques:** Acquired skills in optimizing deep learning models for better performance, learning to adjust hyperparameters and architecture to suit specific types of data.

- **Predictive Analytics Proficiency:** Grew proficient in predictive analytics, understanding how to transform historical data into actionable insights that can significantly impact various sectors dependent on weather conditions.

**Key Insights:**

- **Forecasting Accuracy:** Demonstrated that LSTMs could improve forecasting accuracy by learning from the temporal dependencies present in weather data, providing more reliable predictions than traditional models.

- **Impact on Critical Decision-Making:** Showcased the potential of accurate weather forecasts to enhance decision-making in sectors like agriculture, aviation, and disaster management, where precise weather information is crucial.

- **Future Potential:** Highlighted the adaptability of LSTM networks for other predictive applications, suggesting possibilities for further research in areas such as environmental monitoring and climate analysis.

**Summary:**

The "Enhanced Weather Prediction with LSTM Networks" project represents a significant step forward in meteorological science, employing deep learning to refine how weather forecasts are generated and interpreted. By harnessing the capabilities of LSTM networks, this initiative not only advances my expertise in machine learning but also contributes to more dependable weather prediction tools that can benefit society at large.

Weather Prediction with LSTM Networks


**Introduction:**

In an ambitious solo project, I developed an advanced predictive maintenance model tailored for the manufacturing industry. The aim was to transform traditional maintenance strategies by predicting equipment failures before they occur, using historical sensor data and operational logs. This proactive approach significantly reduces downtime and cuts maintenance costs, marking a substantial step forward in industrial operations.

**What I Did:**

- **Project Leadership:** I single-handedly spearheaded the initiative, from conceptualization to execution, focusing on leveraging machine learning to enhance equipment reliability and maintenance efficiency.
- **Data Preprocessing Mastery:** I meticulously cleaned and transformed extensive operational logs into pristine datasets, creating a robust foundation for model training and analysis.
- **Feature Engineering Expertise:** I employed advanced techniques to extract and engineer key features from complex time-series data, revealing critical patterns and trends indicative of potential equipment failures.

**What I Used:**

- **Modeling Arsenal:** I utilized a combination of Random Forests and Support Vector Machines for precise classification, alongside ARIMA models for time-series forecasting. The implementation was done in Python, using Scikit-learn and PyTorch to ensure optimal performance and scalability.
- **Technological Proficiency:** My choice of technologies and frameworks was driven by their proven track records in handling large-scale data and complex model requirements, ensuring that the project was built on a solid technological base.

**What I Learned:**

- **Deepened Technical Expertise:** This project allowed me to deepen my technical skills in machine learning, especially in the nuanced areas of time-series analysis and predictive modeling within the demanding context of industrial operations.
- **Strategic Problem Solving:** I refined my ability to develop strategic solutions that address real-world problems, such as unplanned downtime and costly repairs, showcasing the practical impact of data-driven decision-making.
- **Independence and Initiative:** Working solo on this project highlighted my capacity for self-direction and my initiative in taking on complex challenges without external guidance, proving my ability to lead and execute significant projects independently.

**Key Insights:**

- **Proactive Maintenance Paradigm:** The project demonstrated the value of shifting from reactive to proactive maintenance strategies, significantly enhancing operational efficiency and reducing costs.
- **Data-Driven Decision Making:** The integration of machine learning with historical data analytics empowers organizations to make informed decisions, preventing equipment failures before they disrupt operations.
- **Continuous Improvement Culture:** The project reinforced the importance of continuous improvement and technological adaptation in maintaining competitiveness in the fast-evolving manufacturing sector.

**Summary:**

Through diligent data preprocessing, sophisticated feature engineering, and the strategic application of machine learning techniques, this solo project not only optimized equipment management processes but also spearheaded a shift towards proactive maintenance strategies in the manufacturing industry. My journey through this initiative not only fortified my expertise in predictive modeling but also ignited a passionate commitment to innovation and excellence in technology-driven industries.


Forecasting Equipment Failures

**Introduction:**

In an innovative approach to classic gaming, I developed an AI-driven project to enhance the gameplay of the iconic Super Mario game. By integrating deep learning techniques, specifically reinforcement learning, Mario is transformed from a character guided by predefined scripts to one that learns and adapts from its environment, promising a dynamic and intelligent gaming experience.

**What I Did:**

- **AI Model Development:** Designed and trained a reinforcement learning model using TensorFlow and Keras. This model enables Mario to understand and interact with the game environment more effectively, learning from each action's outcomes to make smarter decisions over time.

- **Environment Interaction:** Integrated the AI model with the game’s environment, allowing for real-time learning and decision-making. This involved creating a feedback loop where the model continuously updates its strategy based on the game state and outcomes of previous actions.

- **Gameplay Testing:** Conducted extensive testing phases to adjust the difficulty levels and ensure that the AI’s learning curve was aligned with an engaging yet challenging gameplay experience. This helped refine the model to respond to various in-game scenarios effectively.

- **Performance Optimization:** Tuned the model to optimize both speed and accuracy of Mario's in-game decisions, ensuring smooth gameplay and enhancing the player’s experience.

**What I Learned:**

- **Advanced AI Techniques:** Gained in-depth knowledge of reinforcement learning algorithms and their application in real-time environments, enhancing my proficiency in building adaptive AI systems.

- **Integration Skills:** Learned to seamlessly integrate AI with existing software systems, an essential skill for developing modern AI-enhanced applications.

- **Balancing Complexity:** Developed an understanding of balancing computational efficiency with complex AI tasks, ensuring the AI system is both effective and efficient.

**Key Insights:**

- **AI as a Game Changer:** Demonstrated how AI can transform traditional gaming into an interactive and evolving experience, where the game adapts to individual player behaviors.

- **Enhancing User Engagement:** The project illustrated how AI-driven characters can lead to more engaging and immersive gameplay, keeping players challenged and interested.

- **Future of Gaming:** Highlighted the potential for further AI integration in games, suggesting a future where AI not only enhances gameplay but also contributes to game narrative and development.

**Summary:**

The "Deep Learning in Super Mario" project represents a significant leap in game development, employing cutting-edge AI technology to breathe new life into a beloved classic. By equipping Mario with the ability to learn and adapt, the game becomes not just a nostalgic replay but a new frontier in interactive entertainment. This initiative has not only advanced my technical skills but also paved the way for further innovations in the gaming industry, promising exciting possibilities for the future of interactive entertainment.


Deep Learning in Mario


![Machine Learning Visualizers Screenshot](/images/project/mlvisu.png)

The Machine Learning Visualizers project is an interactive showcase of various machine learning algorithms, implemented using Streamlit. This repository hosts a series of applications that demonstrate the inner workings and applications of different algorithms through real-time, interactive visualizations. Users can tweak algorithm parameters on-the-fly and immediately observe the effects on model performance and decision processes.

## Algorithms Covered

- **Naive Bayes Classifier**: [Explore](https://machine-learning-visualizers-bayes.streamlit.app/) the basics of Naive Bayes and its application in classification tasks.
- **K-Means Clustering**: [Understand](https://machine-learning-kmeans.streamlit.app/) how K-Means clustering works to partition data into k clusters.
- **Linear Regression**: [Visualize](https://machine-learning-visualizers-linear-regression.streamlit.app/) the relationship between variables and how Linear Regression can be used to predict outcomes.
- **Random Forest Classifier**: [Learn](https://machine-learning-visualizers-random-forest.streamlit.app/) about the ensemble approach using multiple decision trees to improve predictive accuracy.
- **Support Vector Machine (SVM)**: [Discover](https://machine-learning-visualizers-svm.streamlit.app/) how SVM can be used for both linear and non-linear classification.
- **Principal Component Analysis (PCA)**: [See](https://machine-learning-visualizers-pca.streamlit.app/) how PCA is used to reduce the dimensionality of data while preserving as much variance as possible.
- **Neural Networks**: [Dive](https://machine-learning-visualizers-neural.streamlit.app/) into the workings of Neural Networks and see how they can model complex patterns in data.

**What did I do?**  
I developed a suite of Streamlit applications that provide intuitive and interactive visualizations for several machine learning algorithms. My role encompassed designing the user interface, integrating the machine learning models with Streamlit, and ensuring real-time responsiveness of visualizations. I focused on creating an educational tool that can be used by students and professionals alike to better understand complex algorithms.

**What I used?**

- **[Python](https://www.python.org/):** The primary programming language for backend development, used for its extensive libraries and community support in machine learning.

- **[Streamlit](https://streamlit.io/):** A fast and easy way to create and share beautiful, custom web apps for machine learning and data science. Streamlit was crucial in building interactive UI components that allow real-time interactions with the algorithms.

- **[Scikit-learn](https://scikit-learn.org/):** A machine learning library for Python, used to implement algorithms such as Naive Bayes, K-Means, Random Forest, and SVM. It provides simple and efficient tools for predictive data analysis.

- **[Matplotlib](https://matplotlib.org/)** and **[Seaborn](https://seaborn.pydata.org/):** Visualization libraries used to plot complex graphs and decision boundaries, helping to illustrate the output of various machine learning algorithms.

**What I learned?**  
This project enhanced my abilities in creating interactive applications with Streamlit and deepened my understanding of machine learning algorithms. I improved my skills in data visualization, particularly in dynamically representing algorithm behavior and decision-making processes. Additionally, I learned to effectively communicate complex concepts in an accessible manner, making advanced machine learning techniques understandable and engaging for a broad audience.

Machine Learning Visualizers


Monte Carlo Simulations are powerful tools used to understand the impact of risk and uncertainty in prediction and modeling problems. In this project, I applied the Monte Carlo Simulation technique to perform risk analysis in financial portfolios, providing insights into potential future scenarios and helping stakeholders make informed decisions.

**What did I do?**

- **Simulation Design:** Developed a series of stochastic models to simulate various scenarios in financial markets. This involved generating random samples for market returns based on historical data to predict future outcomes with varying degrees of risk.
  
- **Risk Assessment:** Analyzed the potential for loss in investment portfolios by running thousands of simulated outcomes. This helped in understanding the most likely profit and loss scenarios, enabling a better strategic response to different market conditions.

- **Visualization:** Created visualizations to depict the distribution of possible outcomes, illustrating the risks and probabilities associated with different investment strategies.

**What I used?**

- **[Python](https://www.python.org/):** Used for scripting the simulations, offering flexibility and powerful integration with other libraries.
- **[NumPy](https://numpy.org/):** Essential for numerical operations, used to generate random variables and handle large arrays of data efficiently.
- **[Matplotlib](https://matplotlib.org/) and [Pandas](https://pandas.pydata.org/):** For visualizing the simulation results and handling data structures for analysis.
- **[SciPy](https://www.scipy.org/):** Utilized for more complex statistical operations beyond basic simulations, enhancing the robustness of the analyses.

**Key Insights:**

- Provided a probabilistic view of potential financial outcomes, allowing for better risk management and investment strategy formulation.
- Demonstrated the versatility of Monte Carlo Simulations in quantifying and addressing financial uncertainties.
  
**What I learned?**

This project enhanced my skills in probabilistic simulations and their applications in finance. I learned to integrate various Python libraries to create complex simulations and extract actionable insights from large datasets. The experience also improved my ability to communicate complex statistical concepts and findings effectively through visualizations.


Monte Carlo Simulation for Risk Analysis


![Pokemon Screenshot](/images/project/Pokemon.png)

This project entails the development of a predictive analytics model for a Pokémon dataset. Leveraging Python and AutoML, the project aimed to predict attributes like speed and legendary status with high accuracy. By applying machine learning techniques and utilizing advanced modeling tools, the project sought to uncover intricate relationships among Pokémon generations and optimize predictive accuracy.

**What did I do?**
I spearheaded the development of the predictive analytics model, utilizing Python and AutoML to achieve an 89% accuracy rate in forecasting Pokémon attributes such as speed and legendary status. Machine learning techniques such as Logistic Regression and Decision Tree Classifier were applied to over 25k records, enhancing model performance. Additionally, I employed data visualization tools like Seaborn and Matplotlib for in-depth analysis, leading to improved feature selection and data integrity.

**What I used?**

- **[Python](https://www.python.org/):** Primary programming language used for data preprocessing, model development, and analysis in the Pokémon project. Python's versatility and rich ecosystem of libraries made it an ideal choice for handling complex data and implementing machine learning algorithms.

- **[Pandas](https://pandas.pydata.org/):** Data manipulation library used for cleaning, transforming, and analyzing the Pokémon dataset. Pandas provided powerful tools for data manipulation and exploration, facilitating efficient data preprocessing and analysis.

- **[Scikit-learn](https://scikit-learn.org/):** Machine learning library utilized for implementing predictive analytics algorithms such as logistic regression and decision tree classifier. Scikit-learn provided a wide range of algorithms and tools for model training, evaluation, and optimization.

- **[Seaborn](https://seaborn.pydata.org/) and [Matplotlib](https://matplotlib.org/):** Data visualization libraries used for creating visualizations and gaining insights into Pokémon characteristics. Seaborn and Matplotlib provided flexible and expressive tools for generating various types of plots and graphs, enhancing data exploration and communication.

- **[AutoML](https://en.wikipedia.org/wiki/Automated_machine_learning) (H2O AutoML):** Automated machine learning tool utilized for advanced modeling and optimization of predictive accuracy in the Pokémon project. AutoML algorithms automatically select and optimize machine learning models, streamlining the model development process and improving predictive performance.

- **[Neural Network](https://en.wikipedia.org/wiki/Neural_network):** Neural networks were employed as part of the machine learning pipeline in the Pokémon project. These deep learning models provided a powerful framework for learning complex patterns and relationships in the Pokémon dataset, contributing to the overall predictive accuracy and performance of the analytics model.

**Key Insights:**

- Identified the most significant predictors of Pokémon capabilities.
- Analyzed how different models handle various attributes and conditions.
- Assessed the impact of missing data and outliers on the robustness of predictions.

**What I learned?**
Through this project, I gained hands-on experience in applying machine learning techniques to real-world datasets, particularly in the context of predictive analytics. I enhanced my skills in Python programming, data preprocessing, model development, and evaluation. Additionally, I learned how to leverage AutoML tools like H2O AutoML for automating the machine learning pipeline and optimizing model performance. Overall, this project provided deep insights into Pokémon characteristics and contributed to a broader understanding of data-driven decision-making processes.

Advanced Prediction Model and Analysis


![Web App Screenshot](/images/project/arch.jpg)


**Introduction:**

In this cutting-edge project, I constructed a real-time data streaming and processing architecture designed to empower instant stock market analytics. Utilizing Python, Apache Kafka, and several AWS services, I engineered a solution capable of handling large volumes of data with improved efficiency and speed, thereby enhancing decision-making processes in the financial sector.

**What I Did:**

- **System Architecture and Setup:** Developed a robust data streaming system using Apache Kafka and AWS EC2 to ensure seamless, real-time data flow. This setup allowed for high-throughput and low-latency processing, which is critical in the fast-paced environment of stock trading.

- **Data Storage and Management:** Configured AWS S3 to store the incoming data streams and used AWS Glue to catalog and organize data effectively. This step was crucial in reducing data retrieval times by 40%, thus enabling quicker strategic analysis.

- **Pipeline Optimization:** Optimized the data processing pipeline, written in Python, to handle and analyze up to 5GB of data daily. This included enhancing the reliability and scalability of the system to accommodate real-time analysis needs without performance degradation.

- **Integration and Analytics:** Utilized AWS Athena for querying and performing advanced analytics on the processed data. This integration allowed for the extraction of actionable insights from large datasets quickly and efficiently.

**What I Learned:**

- **Complex Data Stream Handling:** Gained valuable experience in managing complex data streams in a real-time context, learning to navigate challenges associated with high volumes and velocities of financial data.

- **AWS Solutions Proficiency:** Enhanced my skills in utilizing AWS solutions for scalable data engineering tasks, deepening my understanding of cloud services in data-intensive scenarios.

- **Real-time Analytics Impact:** Understood the critical role of real-time analytics in financial decision-making, recognizing how timely data processing can significantly influence trading strategies and outcomes.

**Key Insights:**

- **Efficiency in Data Processing:** Demonstrated the importance of efficient data architecture in reducing latency and increasing the timeliness of insights, which is paramount in stock market analytics.

- **Strategic Advantage through Technology:** Showed how advanced data engineering techniques and tools can provide a strategic advantage by enabling faster, more informed decisions in stock trading.

- **Future Directions in Financial Analytics:** Explored the potential for further innovations in financial technology, particularly in enhancing real-time data processing and analytics capabilities.

**Summary:**

The "Real-Time Stock Market Data Processing" project exemplifies the integration of technology and finance, where sophisticated data engineering solutions directly impact the effectiveness of stock market operations. Through this project, I have not only advanced my technical expertise in data streaming and cloud services but have also contributed to the evolution of data strategies in financial analytics, setting the stage for future technological advancements in the industry.


Real-Time Stock Market Data Processing


**Introduction:**

In the dynamic landscape of social media, understanding public sentiment is crucial for shaping effective marketing and communication strategies. This project focused on analyzing sentiments expressed in social media posts across various platforms to gauge public opinion on different topics. The aim was to equip businesses with actionable insights into consumer perceptions, thereby enhancing their decision-making processes and marketing strategies.

**What I Did:**

- **Data Collection and Cleaning:** Spearheaded the extraction of vast amounts of social media data, implementing rigorous cleaning techniques to ensure the quality and reliability of the data for analysis. This stage was critical in preparing a solid dataset that reflects genuine user sentiments.

- **Natural Language Processing Techniques:** Applied advanced NLP techniques using NLTK and SpaCy for text normalization and tokenization. I developed algorithms to accurately identify and categorize sentiments in unstructured text, which is pivotal in understanding the nuances of public opinion.

- **Model Development:** Engineered and deployed several machine learning models, including Naive Bayes, Logistic Regression, and advanced neural networks like RNNs and LSTMs. These models were crucial in dissecting complex sentence structures and emotional nuances, providing a granular analysis of sentiments.

- **Visualization and Reporting:** Designed and implemented a comprehensive dashboard that visualizes sentiment trends and patterns over time. This tool has been instrumental for stakeholders to visualize data-driven insights, facilitating informed decision-making and strategic planning.

**What I Learned:**

- **Enhanced NLP Skills:** This project significantly advanced my expertise in natural language processing, especially in applying it to diverse and large-scale social media datasets.

- **Machine Learning Application:** I honed my skills in applying various machine learning algorithms to real-world problems, enhancing my ability to tailor models to specific data characteristics and business needs.

- **Data Visualization Proficiency:** Developed a strong acumen in creating intuitive and actionable visualizations for complex datasets, which are essential for conveying findings to non-technical stakeholders.

**Key Insights:**

- **Consumer Insight Through Sentiment:** The project underscored the power of sentiment analysis in uncovering hidden consumer insights, which can dramatically influence marketing strategies and brand positioning.

- **Real-time Market Pulse:** Demonstrated how real-time analysis of social media can serve as a pulse for market trends and public opinion, allowing companies to stay agile and responsive.

**Summary:**

Through meticulous data management, sophisticated natural language processing, and the strategic application of machine learning models, this Social Media Sentiment Analysis project provided deep insights into public sentiment. It has empowered businesses to navigate public opinion more effectively, ensuring that their strategies are aligned with consumer attitudes and market dynamics. My journey through this project has fortified my capabilities in predictive analytics and instilled a robust framework for continuous innovation in data science.

Social Media Sentiment Analysis


The Splitwise Database System project involves creating a sophisticated database model to manage shared expenses effectively. Inspired by the practical utility of the Splitwise app, this database system caters to the everyday financial interactions among users, facilitating seamless and transparent expense sharing.

**What did I do?**  
I designed and implemented a relational database that captures the complexity of shared financial transactions. Key responsibilities included modeling the database schema, enforcing data integrity through normalization up to the Third Normal Form (3NF), and implementing constraints to maintain consistency and reliability of the data.

**What I used?**

- **Database Design Principles:** Employed advanced database design techniques to create an efficient, scalable, and robust system. This involved detailed planning of table structures, relationships, and normalization to ensure data integrity and reduce redundancy.

- **[SQL](https://www.mysql.com/):** Used SQL for scripting database schemas, queries, and for implementing stored procedures, triggers, and views that enhance the functionality and maintainability of the database system.

- **[db4o](http://www.db4o.com/):** Utilized this object-oriented database management system to handle complex data relationships more naturally than traditional relational databases, providing an efficient means of storing objects directly.

**What I learned?**  
This project significantly enhanced my understanding of practical database design and the application of normalization principles. I developed a deeper appreciation for the challenges of managing complex data structures in a real-world application. Additionally, I gained valuable experience in SQL scripting, optimizing database performance, and implementing security measures to protect sensitive data.

Splitwise Database System


**Introduction:**

This project focused on developing a sophisticated machine learning model to predict stock market trends, offering investors and financial analysts a tool to anticipate market movements and enhance their investment strategies. By leveraging historical data and advanced predictive algorithms, this initiative aimed to provide actionable insights that help optimize investment decisions.

**What I Did:**

- **Data Collection and Preprocessing:** Compiled extensive datasets from various financial markets, including stocks, indices, and other relevant economic indicators. The data was cleaned, normalized, and transformed to create a unified structure suitable for analysis.

- **Feature Engineering:** Developed a set of predictive features based on technical indicators, market sentiment analysis, and macroeconomic factors. This step was crucial in distilling complex historical data into meaningful inputs for our models.

- **Model Development:** Utilized advanced deep learning techniques, employing neural networks implemented with TensorFlow and Keras. These models were designed to capture complex patterns and dependencies in market data that are not apparent to human analysts.

- **Backtesting and Optimization:** Performed rigorous backtesting of the models using historical data to evaluate their predictive accuracy and robustness. The models were continuously refined and optimized based on their performance to improve prediction accuracy.

- **Real-time Simulation and Deployment:** Integrated the model into a real-time trading simulation platform to test its effectiveness in live market conditions. This allowed for further refinement of the strategy and adjustments to better capture market dynamics.

**What I Learned:**

- **Advanced Financial Modeling:** Enhanced my understanding of financial markets and how to model them using machine learning. This included gaining insights into the specific challenges of financial data, such as noise, volatility, and non-stationarity.

- **Deep Learning in Finance:** Gained hands-on experience with cutting-edge deep learning frameworks and techniques, learning how to apply these tools to solve complex predictive problems in the finance sector.

- **Strategy Development:** Learned how to translate analytical insights into actionable trading strategies that can dynamically adjust to new information and market conditions.

**Key Insights:**

- **Predictive Power of AI:** Demonstrated the capability of artificial intelligence to predict stock market trends, underscoring the potential of machine learning to transform financial analysis and decision-making.

- **Enhanced Investment Strategies:** Showed how predictive analytics can be integrated into investment strategies to reduce risk and maximize returns, providing a competitive edge in the fast-paced world of finance.

- **Future of Financial Technology:** Highlighted the growing role of machine learning in financial technology, suggesting that the future of finance will heavily rely on predictive analytics to drive decisions.

**Summary:**

The "Machine Learning for Predictive Stock Market Analytics" project represents a breakthrough in financial technology, using artificial intelligence to predict market movements. This initiative not only advanced my skills in machine learning and financial modeling but also contributed to the broader field of financial technology by demonstrating the potential of AI to revolutionize investment strategies and market analysis.

---

Machine Learning for Predictive Stock Market Analytics

![Web App Screenshot](/images/project/wecare.png)

[Case Study PDF](/wecarecasestudy.pdf){:target="_blank"}

[Figma Volunteer Frames Export PDF](volunteer.pdf){:target="_blank"}

[Figma User Frames Export PDF](user.pdf){:target="_blank"}

**Introduction:**

"We Care" is a revolutionary digital platform designed to streamline long-distance caregiving. This project focuses on creating an intuitive and accessible user interface that connects adult children with their aging parents, regardless of geographic barriers. By integrating essential caregiving services into one comprehensive application, "We Care" aims to simplify the management of caregiving tasks while enhancing the quality of life for the elderly.

**What I Did:**

- **User-Centered Design Process:** Implemented a full-cycle design thinking approach, including empathizing with users, defining the problem, ideating solutions, prototyping, and testing. This process ensured that the final product truly resonates with the needs of our users.
  
- **Extensive User Research:** Conducted detailed user research, including creating personas like Kalpana Singh and Arjun Singh to represent our target audience. This research helped in understanding the specific needs, frustrations, and goals of our users, guiding the design process effectively.
  
- **SWOT Analysis:** Performed a SWOT analysis to identify strengths, weaknesses, opportunities, and threats, shaping strategic decisions in the app’s development and feature prioritization.
  
- **High-Fidelity Prototyping:** Utilized Figma and Adobe XD for high-fidelity prototyping, presenting stakeholders with a clear and interactive representation of the user interface, which includes navigation flows and detailed design elements.

**What I Learned:**

- **Advanced Prototyping and Wireframing Skills:** Enhanced my proficiency in using tools like Figma and Adobe XD to create dynamic prototypes and wireframes that effectively communicate design concepts and facilitate user testing.
  
- **Strategic Design Implementation:** Gained insights into strategic planning in design, learning how to align user needs with business objectives using methods like the MoSCoW approach to prioritize features throughout the design and development phases.
  
- **Empathetic User Engagement:** Improved my ability to engage with users empathetically, an essential skill in designing solutions that address real-world problems in caregiving, enhancing user satisfaction and overall usability of the platform.

**Key Insights:**

- **Proactive Solution to Caregiving:** The project underscored the importance of proactive design solutions in addressing the complexities of long-distance caregiving, providing tools that offer peace of mind to families while ensuring the elderly receive the care they need.
  
- **Impact of Intuitive Design on User Adoption:** Demonstrated how intuitive design and thoughtful user interface implementation can significantly affect user adoption and satisfaction, especially among non-tech-savvy elderly users.

**Summary:**

The "We Care" project exemplifies the power of thoughtful UI/UX design in transforming caregiving solutions. Through meticulous user research, empathetic design practices, and strategic feature implementation, the project not only simplifies the caregiving process but also fosters stronger familial connections across distances. This initiative has set a new standard in the caregiving industry, proving that design can be a powerful tool in enhancing lives and solving complex social challenges.



We Care - Long-Distance Caregiving with Intuitive Design

<p>During my internship at Brigham and Women’s Hospital, I worked with the pharmacology team to fix their painfully slow genomic analysis pipeline. Researchers were waiting 2-3 days for results, which was killing their productivity and frustrating everyone involved.</p>
<h2>The Problem</h2>
<p>The team was analyzing 5TB+ of multi-omics data - RNA sequencing, protein expression, metabolomics, plus clinical metadata. Their pipeline was a collection of bash scripts that:</p>
<ul>
<li>Processed everything sequentially (one sample at a time)</li>
<li>Used text files for massive joins</li>
<li>Requested way more cluster resources than needed (500GB memory for 50GB jobs)</li>
<li>Could only run on the ERIS cluster, no local testing possible</li>
</ul>
<p>Between the inefficient code and cluster queue times, researchers waited 48-72 hours just to find out if their analysis worked. If it failed, that’s another 2-3 days down the drain.</p>
<h2>What I Did</h2>
<p>First, I profiled the existing pipeline to understand where time was being wasted. Turns out, most of the “computation time” was actually:</p>
<ul>
<li>Waiting in queue due to overestimated resource requests</li>
<li>Reading/writing massive text files repeatedly</li>
<li>Processing samples one by one instead of in parallel</li>
</ul>
<p>I rebuilt the pipeline using PySpark and switched the storage format to Parquet:</p>
<ul>
<li><strong>Parquet files</strong>: 60% smaller than text files, way faster for column-based queries</li>
<li><strong>Parallel processing</strong>: Partitioned by sample ID so multiple samples process simultaneously</li>
<li><strong>Right-sized resources</strong>: Profiled actual memory usage and requested appropriate amounts</li>
<li><strong>Local testing</strong>: Set up Jupyter notebooks so researchers could test locally before cluster submission</li>
</ul>
<h2>The Results</h2>
<p>The improvements were dramatic:</p>
<ul>
<li>Processing time: 48-72 hours → 15 minutes</li>
<li>Queue wait time: 2-6 hours → 5-10 minutes</li>
<li>Storage needs: Reduced by 60%</li>
<li>Researcher productivity: From 2-3 analyses per week to 10-15</li>
</ul>
<p>The 15-minute runtime includes all the joins, transformations, and statistical calculations. The speedup came from:</p>
<ul>
<li>Parallel processing (biggest win)</li>
<li>Columnar storage format (Parquet)</li>
<li>Eliminating redundant I/O operations</li>
<li>Actually using the cluster’s capabilities</li>
</ul>
<h2>What Made This Work</h2>
<p>The key insight was that this wasn’t really a “big data” problem - it was a “badly organized data” problem. The cluster had plenty of power; the pipeline just wasn’t using it properly.</p>
<p>I also learned that scientists don’t care about fancy features - they want reliability and speed. So I kept the interface simple, wrote clear documentation, and made sure they could modify the pipeline themselves after I left.</p>
<h2>Technical Details</h2>
<ul>
<li><strong>Data</strong>: 1.3TB of genomic and clinical data</li>
<li><strong>Cluster</strong>: Harvard Medical School’s ERIS HPC cluster</li>
<li><strong>Old stack</strong>: Bash scripts + text files + sequential processing</li>
<li><strong>New stack</strong>: PySpark + Parquet + parallel processing + Jupyter</li>
<li><strong>Optimization</strong>: Proper resource allocation, partition strategy, columnar storage</li>
</ul>
<p>By the end of my internship, the team could iterate on analyses in real-time instead of waiting days. They were exploring hypotheses they’d previously avoided because the computational cost was too high.</p>
<p><em>Note: Specific code and implementation details remain confidential per BWH/Harvard Medical School policies.</em></p>
