Projects

Multimodal meeting copilot that watches shared screens, listens to audio, powers contextual chat, runs fact-check workflows, and generates final summaries. Built at the Multimodal Frontier Hackathon in San Francisco and awarded Best Use of Railtracks.

Retrieval-Augmented Generation system built for EECS coursework and research applications.

Studying Pareto frontier between user engagement and lifecycle carbon footprint in large-scale recommendation systems. BPR and SASRec models with carbon-aware re-ranking.

Deep research agent: layered search and extraction, live knowledge graph while the run is active, Learning trail for strategy shifts, and markdown reports. Built and presented at AWS Builder Loft (San Francisco). Stack includes Next.js, FastAPI, Claude, Tavily, Bedrock, Aerospike, and Airbyte.

Open-source benchmark harness for comparing LLM inference performance across providers and configurations. Latest vLLM finding (single-worker local setup, Qwen2-0.5B-Instruct): higher concurrency mostly increased queueing latency rather than throughput; doubling max_tokens roughly doubled latency while tokens/sec stayed similar, suggesting generation rate as the bottleneck.

Got tired of every PR review tool requiring account sign-ups and complex setup, so I built my own. Drop-in GitHub Action that runs LLM-powered code review with structured, inline comments — severity, category, confidence, and suggested fixes. Supports single-pass and multi-pass modes (correctness, security, performance). Also works as a standalone CLI. Published on the GitHub Marketplace.

Built an automated LLM-based testing agent to fuzz and validate AI-generated web applications.

Multi-stage LLM agent pipeline to parse freelance job ads, assess feasibility/margins, generate bids, execution plans, and autonomously produce deliverables.

MuZero reinforcement learning agent for Snake. Full training pipeline including MCTS, neural network policy/value estimation, and evaluation framework.

ML models to predict maritime vessel trajectories using historical AIS data. Feature engineering pipeline and evaluation across multiple modeling approaches.