I've spent 10+ years building frontend systems — from component libraries used by thousands of engineers to real-time data dashboards that handle millions of events. But somewhere along the way, the tooling shifted, and I went all-in on AI engineering. Not prompt tweaking, but the unglamorous infrastructure work: retrieval pipelines, structured extraction with BAML, stateful agent workflows in LangGraph, and the operational layer that keeps LLMs actually useful in production.
Right now I'm building BIOS — an AI-native health platform that turns fragmented personal health data into something you can actually reason with. The full stack: React Native (Expo) on the client, TanStack Start + Hono on the edge, Cloudflare Workers + D1 for globally distributed persistence, and a RAG layer backed by pgvector. v1 is live. I also run Tea Party, a human-AI co-creation environment where I've been designing multi-agent memory systems and async agent coordination — part research, part daily workflow.
I care about depth over breadth. I'd rather build one system properly — with a coherent data model, clear failure modes, and architecture that can grow — than ship five things that fall apart under pressure. If you're building something that sits at the intersection of AI and real user workflows, let's talk.
A conversation-driven co-creation space — my daily environment for building with AI. Traditional AI assistants are stateless; every session starts from zero. I designed a 4-layer memory architecture with tiered context loading to keep agents coherent across sessions, then layered in multi-agent coordination, async task delivery, and an Idea → Research → Launch lifecycle. Every project I ship, including this portfolio, originates here.
An AI-native health companion that turns fragmented personal data into something AI can actually reason about. Existing apps are either cold data loggers or annoying habit trackers — none truly know your body. Gemini Flash handles food recognition and intent parsing; Claude Sonnet generates weekly narrative reports. Shipped v1 with a Skill architecture, a Credits economy model, and a data fingerprint cache that significantly cuts AI call costs.
A 1,000-faces interview prep platform — every user sees a completely different experience. Existing platforms serve static question banks that can't adapt to individual backgrounds or target roles. Users describe their background and goal; AI dynamically generates a personalized knowledge map, study materials, and mock interviews on demand. Built on Cloudflare Workers AI with BAML-structured output for quality and consistency. POC live with 10 courses and 181 practice problems.
I build things that sit at the intersection of AI and real user workflows. If you're working on something like that — or just want to talk through an idea — I'd genuinely like to hear from you.
Say Hello hello@example.com GitHub LinkedIn