Overview
RAG systems, reinforcement learning, full-stack, and data.
The things I've built. Start with the flagship, a game with a published paper behind it, or follow any card straight to its code.
An adversarial dungeon game where a PPO-trained agent hunts you through fog of war. Playable here, and the subject of a published paper.
A production pipeline at work that turns plain-English questions over multilingual survey data into guarded, read-only SQL and narrates the exact result.
A production RAG assistant that grounds each student in their own classes: the recordings themselves plus the material the class was taught from. It handles student audio, so the deep write-up is shared on request.
An online entrance-exam portal with in-browser proctoring, used for real admission intakes at work.
Term-long college research: we took SmokeyNet, a published wildfire smoke-detection model, got its codebase training on our own hardware, and tested our own improvement ideas against it, from background subtraction to captioning to anomaly detection. Clouds defeated most of them, and learning why was the point.
A retrieval-augmented chatbot over a mental-health corpus, returning grounded answers with citations.
Document-intelligence RAG built for a Cognizant hackathon: two pipelines built in parallel, one fully local and one API-based, then benchmarked head to head. The local build won on context recall and answer faithfulness. My slice was the model integration and the benchmarking.
My 10th grade final project, a console Java ATM that kept its records in a CSV file, redesigned in my free time as a full-stack app with an actual UI. A homage, not a flagship.
Trending-video analysis across the US and India, delivered as an interactive dashboard.