Projects

A collection of my work in AI, machine learning, and technology.

AI Agent

Kaggie

Kaggie is an AI-powered Chrome extension that serves as your intelligent Kaggle competition assistant. It leverages multi-agent LangGraph systems with adaptive RAG to provide personalized strategies, insights, and guidance directly within your browser while competing in machine learning challenges.

PythonLangchainLanggraphFastAPIPineconeOpenAI APIReactTypeScript

Built Chrome extension with LangGraph multi-agent system using OpenAI API and adaptive RAG pipeline.

Automated Kaggle scraper with Playwright and spaCy NLP, deployed via GitHub Actions every 4 hours.

Engineered Pinecone vector database with semantic chunking and Firestore filtering for qualified discussions.

Designed adaptive RAG evaluation system with LLM scoring and intelligent retry/fallback strategies.

Built FastAPI backend with vector search and competition data retrieval endpoints.

Created React TypeScript frontend with Chrome storage persistence and real-time streaming responses.

Built conversation summarization with summarization agent optimizing long conversations.

AI Agent

CLARA AI

Clara AI is your AI-powered executive assistant for email. It leverages advanced language models to triage, classify, and draft responses to emails, integrating seamlessly with Gmail.

PythonLangchainLanggraphGmail APIAzure OpenAIAzure CosmosDB

Developed an agent-based AI assistant that automates Gmail triage and response drafting using LLMs and workflow orchestration.

Engineered agent workflows with LangChain and LangGraph for context-aware, multi-step email automation.

Integrated OAuth2, Gmail API, and Azure CosmosDB for secure, persistent, and seamless user experience.

Built a CLI for real-time status and user interaction.

Time-Series Forecasting

BNB Horizon

A sophisticated deep learning system for cryptocurrency price prediction, specifically focused on Binance Coin (BNB). The project combines advanced time series analysis with real-time market data to provide price forecasts at 5-minute intervals.

PythonPyTorchPandasHugging Face TransformersPostgreSQLFlaskReactNode.js

Fine-tuned IBM's Granite-TimeSeries-TTM-R2 deep learning model for precise BNB price forecasting

Implemented a robust data pipeline for processing and analyzing 5-minute interval kline data

Designed and optimized PostgreSQL database schema for efficient storage of historical data and predictions

Developed a Flask REST API with comprehensive error handling and rate limiting

Created CI/CD pipelines for automated model deployment and system updates

Image Classification

OASIS Alzheimer Classifier

An innovative deep learning solution for early-stage Alzheimer's disease detection using brain MRI scans. The system employs a lightweight CNN architecture to provide rapid and accurate classification of disease stages, assisting medical professionals in early diagnosis.

PythonPyTorchOpenCVTensorFlow

Architected and implemented a custom TinyVGG CNN with 23k parameters from scratch

Processed and analyzed 86k OASIS MRI samples for 4-class disease stage classification

Achieved 96% accuracy through optimized model architecture and training pipeline

Implemented weighted loss functions to address class imbalance in the dataset

Developed data augmentation techniques to enhance model robustness

Achieved balanced F1-score of 96% across all disease stages

Continous Control Tasks

SatNet: Skeletal Attention Network

A cutting-edge reinforcement learning system for articulated robot control, combining Graph Attention Networks (GAT) with Beta distribution for precise and reliable motion planning. The project focuses on improving sample efficiency and robustness in continuous control tasks.

PythonPyTorchGymnasiumStable-baselines3PandasNumpy

Designed and implemented a novel GAT-based architecture for robot control

Integrated Beta distribution for improved action space exploration

Developed custom attention mechanisms for better spatial understanding

Implemented efficient sampling strategies for enhanced learning speed

Created comprehensive testing framework for control stability

Achieved significant improvements in sample efficiency and task completion rates

NLP

Amazon Reviews Sentiment Analysis

A comprehensive sentiment analysis system for Amazon product reviews, processing millions of customer feedback entries to provide actionable insights for businesses. The project demonstrates advanced NLP techniques and efficient data processing at scale.

PythonNLTKScikit-learnPandas

Developed and optimized a Logistic Regression model for sentiment classification

Processed and analyzed 3.6M Amazon reviews with balanced class distribution

Achieved 92% accuracy through advanced feature engineering techniques

Implemented efficient text preprocessing pipeline for large-scale data

Created comprehensive documentation and methodology guide

Published detailed analysis and findings in a Medium article

NLP

Shakespeare-GPT

An innovative language model project that recreates Shakespeare's writing style using a custom Transformer architecture. The project demonstrates deep understanding of transformer models and efficient training techniques for creative text generation.

PythonPyTorchTransformersHugging Face

Architected and implemented a custom NanoGPT model with 9M parameters

Developed specialized character-based tokenization for Shakespearean text

Optimized training pipeline for efficient convergence on limited hardware

Implemented custom batching strategies for improved training efficiency

Created evaluation metrics for style matching accuracy

Achieved high-quality text generation with authentic Shakespearean style