I build AI systems that hold up in production.

Problem

Users ask health questions that have to be answered from their own data, meaning their wearable signals, their logs, and their history, with no leakage between users, and it has to stream fast enough to feel like a conversation. At around 9 seconds to an answer, people stopped engaging with it.

Approach

A TypeScript retrieval stack on Supabase Edge Functions: hybrid dense + BM25 retrieval, Voyage cross-encoder reranking, streamed Claude responses, and a pgvector semantic cache scoped per user with dedicated safety guards.

Problem

Testing how an autonomous vehicle perceives edge cases usually means capturing those situations in the real world, which is expensive, slow, and in some cases dangerous. The scenarios you most need to test, like a pedestrian at night or heavy glare or an occluded object, are exactly the ones you cannot arrange whenever you want.

Approach

I built a pipeline that chains four AI systems together. GPT-4o takes a one line prompt and expands it into three scenario variants that each target a different perception challenge. World Labs Marble turns those into navigable 3D worlds. YOLOv8 runs object detection on the generated scenes. Then GPT-4o reads the detection results and produces a Perception Difficulty Score from 0 to 100, along with a risk level and recommendations.

Problem

A 90 minute lecture recording is very hard to search through, and Panopto captions sit behind a session cookie, so nothing outside the browser can reach them. There is also a harder question underneath that one. Once I had hand built a multi agent pipeline for taking notes, I had no real way of knowing whether the prompts and structure I chose were any good.

Approach

PanNote is a Chrome MV3 extension that detects a Panopto lecture, pulls the captions in the page itself, and runs the transcript through a multi agent GPT-4.1 Mini pipeline that renders structured notes in a side panel. PipelineEvolve treats that pipeline configuration, meaning the prompts and the agent topology, as a program that can be optimized. It searches for a better configuration using LLM driven mutation and a locally trained proposer, and scores every candidate against the pipeline I built by hand.

Problem

Legal documents break simple RAG setups. Clauses reference each other, definitions sit pages away from where they are used, and a wrong answer is worse than no answer at all. Standard top-k retrieval kept missing the clause that actually mattered. Once I fixed retrieval, a stock 3B model still could not read a clause and answer the question correctly.

Approach

I built hierarchical FAISS retrieval that goes from document to section to chunk, added BM25 hybrid recall and cross-encoder reranking, routed generation across OpenAI, Anthropic, and a local llama.cpp model, and fine-tuned Qwen2.5-3B with LoRA on 961 lease question and answer pairs that I built from real lease clauses. All of it is graded by an evaluation harness that I wrote before any of the optimization.