Introduction
Building production RAG systems, reinforcement-learning agents, and applied NLP.
AI/ML engineer. I build RAG systems, train reinforcement-learning agents, and ship NLP.
A bias I'll admit to: most of what ships as "AI" is retrieval with a good system prompt, and benchmark leaderboards are mostly marketing. What actually matters is whether a system stays reliable on real inputs and honest about what it doesn't know. That is the bar I hold my work to.
I'm an AI & ML Associate at Impact Solutions Lab, an AAM Foundation initiative, working on the Program and Curriculum vertical. Before this I interned here and on a couple of ML and analytics teams. Somewhere in between, I trained an agent to hunt people through a dungeon and wrote a paper about it.
What I work on
- Retrieval-augmented generation: chatbots and document-intelligence systems that answer from real sources, with citations instead of guesses. The AI twin on this site is one; a mental-health assistant and a hackathon-built document RAG are others.
- Reinforcement learning: agents that behave, not just score. Hunter Wumpus is the flagship, a PPO agent that hunts the player from memory and the subject of a published paper.
- NLP and classical ML: attention models, forecasting, and the unglamorous end-to-end work from data cleaning to a model that actually ships.
- The glue around it: full-stack apps in React, Node, and FastAPI, plus automation on GitHub Actions and the dashboards that make results legible.
Right now
Full-time at Impact Solutions Lab, I build AI systems for an education nonprofit. Three of them are live: a survey-intelligence pipeline that answers plain-English questions with guarded, read-only SQL, a learning assistant that grounds each student in their own classes and the material they were taught from, and the internal exam platform the school runs its own term exams on. The internal tooling I built as an intern, the Drive-to-YouTube pipeline and the entrance-exam portal, still runs in production.
About this site
This site is my documentation. Everything here is versioned, tested in production, and occasionally hunted by a Wumpus. Read the project write-ups, put my AI twin to the test, or play the reinforcement-learning game the paper is about.