I build software
people actually use.

I wrote a Gymnasium-style simulation of the game plus a bruteforce expectimax solver to act as an expert, then trained a Decision Transformer dealer policy (behavior-cloning the solver, with return-conditioning for a difficulty dial). Training ran on Modal GPUs with batched data collection, async evaluation, and torch.compile for roughly 4–5× faster iterations.

The interesting part was the debugging: an early model collapsed to always shooting the player. I traced it to environment bugs poisoning the training labels and to the sequence model overfitting its context window — fixing both took it from ~22% to ~46% win rate against the optimal solver (and ~86% vs a random opponent). I then decompiled the real Godot game and wrote a mod that hooks the dealer's decision and calls a local Python inference server, so my trained agent actually plays inside the game.

A personal project, end to end: simulation, an expert solver, offline-RL training in the cloud, hard debugging, and real-game integration via decomp + modding. In my testing the final policy tracks near-optimal play. Happy to walk through the training and the in-game bridge.

Every day, staff have to decide which CEAs go to which classrooms — and that shifts constantly as students and staff come and go. I built CEA Assigner so they can produce a sensible assignment in seconds instead of working it out by hand.

I chose Apps Script deliberately: it lives inside the Google Workspace tools the school already uses, needs no extra logins or infrastructure, and stays inside the school's existing data boundary — which matters for a tool touching student and staff information.

In active daily use by school staff — it removes a recurring manual scheduling headache for the people who run the building.

You talk, the robot acts. "Move forward a good bit," "start spinning around," "what does the sign say?" — each spoken command is transcribed, handed to an LLM that decides what the robot should do, and turned into drive commands the robot carries out. It isn't driving blind, either: ask it about a printed card held in front of it and it reads the text through its camera and answers.

I built the bridge between the language model and the real hardware. A Jetson Nano I added does the onboard work — capturing voice and camera and running the control loop — an LLM interprets each request against what the cameras see, and its decisions are passed down to the RoboRIO, the FRC controller already on the robot, which handles the low-level motor control. No new drivetrain, no rebuild — I gave the robot we already had eyes, ears, and an agent in the loop.

My high-school capstone — a working prototype, demoed live in the shop, where my interest in agentic AI meets real hardware: voice in, an LLM reasoning over what it sees, motors out. It's a rough-and-ready demo, not a polished product, and I'm happy to walk through the architecture or show more footage.