I ship AI agentsand the appsthey live in.

The single AI orchestrator behind Fig's one-chat-box product. A LangGraph plan-and-perform loop on FastAPI + Claude drives every turn — building and deploying websites, generating image/video/music/slide media, researching the web and acting on the user's connected apps, all in one graph instead of a separate planner and executor. Parallelizable work fans out to recursive subagents: the same graph re-run with a narrowed toolset and a specialized prompt, so deep research can hit ~15 sources at once with no separate worker service. Turns stream over a resumable SSE block protocol backed by Redis event streams and PostgreSQL checkpointers, and a modular skills framework plus MCP servers load capabilities on demand.

• LangGraph
• FastAPI
• Claude
• MCP
• Subagents
• Python
• SSE
• Redis
• PostgreSQL

A node-based creative studio: text, image, video and audio models wired together on an infinite canvas, with budgets, approvals and brand memory around every paid generation. I'm a core contributor — 440+ commits across ROOTS, the agent that reads the whole canvas graph and builds workflows on it (typed skills like character and product sheets, plan repair, memory), and the Video Studio I built end to end: a reference video cut into shots, clip units and contact sheets in one ffmpeg pass, storyboard frames generated through the canonical path, clips shot motion-first or frame-first across Higgsfield, fal and OpenRouter, and the final cut assembled from a frozen manifest. Plus model routing and pricing across providers. React 18, Fastify 5, Neon Postgres, BullMQ and a FastAPI + LangGraph agent service.

• React
• Fastify
• LangGraph
• FastAPI
• Postgres
• BullMQ
• ffmpeg
• OpenRouter
• fal
• Agents

A generalist agent that works inside a real Chrome over the DevTools Protocol, using structural page understanding instead of a screenshot loop: one observation is a ref-annotated accessibility tree computed in a single DOM pass — roughly 120 to 400 tokens and 1 to 2 ms — that the model acts on by reference. It runs against a dedicated Chrome profile it owns, so it never touches your everyday browsing. Every outbound action is approval-gated and journaled, so a failure mid-send is never narrated as done or silently retried; contacts are deduplicated in the database, not by similarity; and a learning cycle turns a real failure into a candidate lesson that a regression gate must pass before promotion. Zero npm dependencies: Node 22 native TypeScript and node:sqlite, with 600+ unit tests and a real-Chrome suite.

• TypeScript
• Node 22
• Chrome DevTools Protocol
• Agents
• SQLite
• Composio
• OpenRouter
• Evals

An end-to-end suite that runs a real business top to bottom — a marketing website, a billing & invoicing / POS app with PDF receipts and QR payments, and an analytics admin dashboard with live charts and Excel/PDF exports. One shared design system, auth and Firebase data layer across all three apps.

• Next.js
• React
• TypeScript
• Firebase
• Tailwind
• Recharts

An AI assistant that runs entirely on your Mac or iPhone. No internet, no API bill, and nothing you type ever leaves the device — which matters most for the personal writing people paste into chatbots without thinking. A Swift SDK runs llama.cpp in-process, picks a model that fits the hardware it finds, and calls the cloud only when a request is genuinely too slow or too battery-expensive to answer locally. Built measurement-first: a benchmark harness produced 52 documented findings across three real devices, including the ones that disproved my own earlier conclusions.

• Swift
• llama.cpp
• iOS
• Python
• On-device AI
• Benchmarking

A secure AI refund agent where a deterministic policy engine enforces authorization in code rather than in prompts — so the model can reason about a case, but it cannot be talked into a payout. Every refund decision passes through typed policy checks before money moves, which makes the whole flow resistant to prompt injection. Python and FastAPI on the back, LangGraph and Claude for the agent loop, React and Vite in front.

• Python
• FastAPI
• LangGraph
• Claude
• React
• Vite
• Security

A consent-aware trust and safety layer for intimate chat platforms. A two-layer pipeline scores every message against a per-conversation consent state machine on-device, and escalates only the genuinely ambiguous cases to Claude — so the common path is fast, private and cheap, and the model is reserved for the judgement calls. TypeScript end to end: Node.js and Express behind, Next.js in front.

• TypeScript
• Node.js
• Express
• Next.js
• Claude
• Trust & Safety