Projects

• Walkthru

  2026

  A pre-commit gate that makes you understand your own diff.

  Git hooks, LLM quiz generation

  Details ↗

  Problem

  AI can write your code faster than you can understand it. Teams ship diffs nobody on the team can explain.

  What I built

  A git pre-commit hook that quizzes the author on their own diff (multiple choice, matching, or free response) and blocks the commit below a 3.0 out of 5 score.

  Outcome

  1st place at the NY Tech Week Intern Hackathon 2026 (Mantle, YC F25).

• Dextrivia

  2026

  Orbital debris removal, treated as an optimization problem.

  Python, SGP4, Celestrak TLE, QUBO

  GitHub ↗

  Details ↗

  Problem

  Removing debris from low Earth orbit is a sequencing problem: N objects, a delta-v cost between every pair, one best visiting order. Structurally, it's the traveling salesman problem.

  What I built

  A pipeline that pulls live TLE data from Celestrak, propagates each object to a common epoch with SGP4, builds a Hohmann-transfer delta-v cost matrix, and solves a greedy baseline sequence for the Iridium-Cosmos 2009 debris cloud.

  Outcome

  The classical benchmark a QUBO solver is being built to beat.

• prism

  2026

  Self-hosted AI code review for GitHub pull requests.

  Python, Ollama, GitHub webhooks

  GitHub ↗

  Details ↗

  Problem

  Cloud review bots send your diff to someone else's API and bury pull requests in nitpicks.

  What I built

  A webhook server that pulls the diff on every PR event, reviews each hunk with a local Ollama model, and posts inline comments on the right file and line. When there's nothing worth saying, it says nothing.

  Outcome

  A review loop where no code leaves the machine.

• The Last Alibi

  2026

  An AI murder mystery where every case is new.

  Next.js, TypeScript, Tailwind, Framer Motion, OpenAI API, Zustand

  GitHub ↗

  Details ↗

  Problem

  Detective games replay the same scripted case. Once you know the killer, the game is over.

  What I built

  A case generator (victim, suspects, motives, timeline, hidden killer), interrogations where the killer lies and the innocent don't, an evidence board that tracks contradictions, and an AI-judged final accusation.

  Outcome

  A complete noir game with difficulty tiers and keyboard shortcuts.

• ragit

  2026

  Chat with any folder of documents, with no API keys and no cloud.

  Python, Ollama, Local embeddings

  GitHub ↗

  Details ↗

  Problem

  Most RAG tools want a cloud vector database and a paid API before you can ask your own documents a question.

  What I built

  A local CLI: `ragit index ./docs`, then `ragit chat ./docs`. Embeddings and generation both run through Ollama on your machine.

  Outcome

  An installable Python package with no external services.

• Orphy

  2026

  A real-time AI vocal coach.

  Gemini API, Next.js, LangChain, Pinecone, SendBlue

  Details ↗

  Problem

  Vocal feedback usually arrives after the performance, when a coach reviews a recording.

  What I built

  An agent that listens live, analyzes pitch and delivery, and coaches you while you sing or speak.

  Outcome

  Built end to end at Multimodal Hacks during NY Tech Week.