Projects

ForeSight: Calibrated Decision Models as Opponent Forecasters

Submitted to AAAI '27

  • Built a Pokemon game agent using a system one opponent forecaster, matching published LLM agents win-rate (56%)
  • Fine-tuned a 421M System One model (Laya) to forecast opponent moves, lifting prediction accuracy from 43% to 80%
  • Showed 2-turn lookahead adds 13–20 pts to win rate over one-step play, using 4% of per-turn budget for lookahead search

Attributing Tool Failures in Bulk Agent Operations

Submitted to NAACL '27

  • Benchmarked LLM agents on AppWorld, finding they misreport which tool call failed in long runs 67% of the time
  • Built a drop-in MCP proxy that echoes call arguments into results, restoring 94–100% accuracy across Claude & GPT

Pass@k is Not a Property of a Model

Python, PyTorch, verl, AWS, SkyPilot

  • Published first-author preprint on evaluating RL-trained LLMs & methods for training/testing; submitted to conference
  • Showed central results of Limit of RLVR (Yue et al.) fall 3x (36 to 12 pts) by altering the wording of the test questions
  • Evaluated that an ensemble of 3 independently trained Qwen2.5 runs beats one model by 6% at matched sample budget
  • Showed the main reasoning benchmark (MATH-500) is 98.8% solved by base models it tests, leaving 6/500 trials usable

Agentic Polymarket Predictor

Python, Neo4j, Qdrant

  • Built a 5-stage agentic predictor for Polymarket, achieving 17% ROI during forward testing & 30.5% ROI for backtesting
  • Architected a hybrid RAG pipeline over Neo4j & Qdrant via Reciprocal Rank Fusion with a BGE reranker
  • Beat or matched the market on Brier scores on the 53% of markets the model chose, validated across multiple evaluations

First Person Boxing Game

Python, PyTorch, ONNX, MediaPipe, Rust, TypeScript, Docker

  • Built a 2 player boxing game engine that tracks movement through your camera and simulates a boxing match, like a Wii
  • Designed an ML model to classify punches into 5 types (ex. jab, hook) at 80% accuracy & compressed it to run in browser
  • Architected the game engine using 33 tracked body points per player 60 times a second — works out from joint speed whether a hit landed and where, then runs damage with hit thresholds calibrated to each player's own reach and speed

CatanRL

Python, PyTorch Geometric, Gymnasium, FastAPI, AWS ECS

  • Built a production multi-agent RL system for Settlers of Catan using MAPPO with self-play, served via FastAPI on AWS ECS Fargate
  • Achieved a 74% win rate vs. random and 52% vs. rule-based agents (25% baseline) with a heterogeneous Graph Attention Network over a typed board graph
  • Built an autonomous research agent using UCB1 bandit selection to propose & promote hyperparameter configs, closing the loop with drift-triggered retraining

Mass-Sender

JavaScript, Gmail/Outlook OAuth2

  • Built an open-source bulk email sender with Gmail and Outlook OAuth2, CSV/XLSX input, rate limiting, and dry-run previews
  • 13 stars on GitHub