Things I've built and learned from along the way.
Won 1st place ($10K) at Breaking Barriers Hackathon sponsored by AWS, NVIDIA, and Anthropic, competing against 40+ teams. Built a personalized learning platform using Claude Haiku to generate tailored content in 4+ formats: narratives, summaries, diagrams, and interactive games.
AI agent observability platform for tracking LLM calls, tool invocations, and decision paths. Features a Python SDK with LangGraph/LangChain adapters, real-time WebSocket streaming, and D3.js DAG visualizations. Includes training data curation with 5-tier quality scoring and embedding clustering. Supports token tracking and cost attribution across 9+ LLM models.
Hybrid ViT-Mamba architecture combining Vision Transformers with State Space Models for efficient gesture recognition. Achieves 72% Top-1 accuracy on Jester dataset (27 classes) with linear time complexity. Features a low-latency pipeline (~16ms latency, 60+ FPS) optimized for real-time applications.
Building a custom coding agent inspired by Claude Code and Q CLI to avoid subscription costs. Leveraging free LLM inference APIs like NVIDIA NIM for intelligent code generation and assistance. Early-stage development focused on core agent architecture and API integration.
Interactive resume analyzer providing targeted feedback. Trained a Logistic Regression model on 2,400+ resumes. Built an NLP pipeline with NLTK and LanguageTool to calculate grammar scores and assess resume metrics.
A full-stack recipe search application that lets users discover recipes by ingredients or dish name using the Edamam Recipe API. Features detailed recipe information, ingredient lists, and a responsive design for seamless browsing across devices.
A Chrome extension that generates concise summaries of YouTube video transcripts. Leverages the YouTube Transcript API for extraction and Hugging Face Transformers for advanced text summarization, with a Flask backend processing requests seamlessly.