Building ML systems and full-stack products. First-author IEEE Xplore publication. Ships things people actually use.
Cornell CS · Cornell Data & Strategy · Gen AI @ Cornell
- Technology Implementation Associate @ Cornell Data & Strategy — LightGBM classifier disambiguating 37.6% of unresolved payer records across 100k+ discharge records; gradient boosting regressor R²=0.90 on 210k SPARCS records
- Developer @ Generative AI at Cornell — 10k+ startup leads pipeline from 5 sources, Claude API cold emails, PostgreSQL + Redis/BullMQ
- Software Engineer Intern @ VHB — PDF-to-form pipeline for 10k+ crash reports/year, RAG-style LLM classification at ~90% accuracy
| Project | What it does | Stack |
|---|---|---|
| Cassandra | 7-stage agentic Polymarket predictor — 87% accuracy, 45% ROI, hybrid RAG over Neo4j + Qdrant with BGE reranker, 96% hallucination catch rate | Python FastAPI Neo4j Qdrant Claude API Modal |
| CatanRL | RL agent for Catan — 74% win rate vs random, 52% vs rule-based; UCB1 bandit for autonomous hyperparameter search; AWS ECS retraining pipeline | Python PyTorch Geometric FastAPI AWS Modal |
| ShadowStrike | Real-time multiplayer fighting game in Rust/WASM — rollback netcode, fixed-point arithmetic, 304-byte snapshots, 50ms P95 latency | Rust WebAssembly WebSockets Canvas2D |
| Legion | Distributed AI agent orchestration in Go — exactly-once task execution via PostgreSQL advisory locks, embedded SFTP server, sub-8µs WebSocket fan-out | Go PostgreSQL WebSockets React |
| FlavorNet | GNN for ingredient compatibility — 436k pairs, heterogeneous GAT, InfoNCE loss, active learning loop, deployed on Vercel + Modal | Python PyTorch Next.js Claude API |
Languages
ML / AI
Web & Infra
