Selected work.

I build CGCircuit with a team of AI agents rather than one model doing everything. Each agent gets an isolated context and a narrow job; an orchestrator owns the requirements and the whole. I call it harness engineering, and I gave a talk on it at Blizzard Cinematics.

• A team, not a monolith. Domain supervisors coordinate specialists — architect, researcher, backend, frontend, security, reviewer — each seeing only what its job needs.
• A disciplined loop. Research the scope, spawn the right agents, build and test in parallel, review from several angles, iterate until the supervisors agree, ship, then monitor after merge.
• Isolated contexts, in parallel. A single agent is a single-core CPU; the harness is a multithreaded machine — separate contexts, work in parallel, a scheduler that owns the whole.
• The harness is the real work. Reproducible, reviewable, maintainable — the opposite of one-prompt “vibe coding” that falls apart the moment scope grows.

An AI instructor built into CGCircuit's courses, so learners get help the moment they need it instead of scrubbing through hours of video for one answer.

• Answers in context. Learners ask questions against the course material and get sent to the exact moment that answers them.
• Search across the library. AI search over lessons and transcripts turns a video catalog into something you can query directly.
• Localized subtitles. Automatic subtitles in multiple languages widen who the courses can reach.

CGCircuit is a learning platform I designed, built, and still run — every layer of it. Storefront, course creation, and a custom video player on the front; a serverless back end, a video pipeline, and the AWS operations behind it.

• Front end. A React storefront, creator tools for authors, and a custom video player.
• Serverless back end. A Python (Chalice) API over DynamoDB with real-time WebSocket connections — auth, integrations, and monitoring included.
• Video pipeline. Automated encoding, transcoding, and optimization, so raw uploads become smooth, adaptive playback.
• Run, not just built. Scaling, cost, analytics, and day-to-day AWS operations — I keep it running long after launch.

A workflow that turns messy input — raw meeting notes — into tracked execution: action items, a planned week, and a clear view of what got done.

• Notes in, actions out. Meeting notes are read automatically and turned into concrete, assignable to-dos.
• A planned week. Those to-dos are scheduled and prioritized into a realistic week rather than a growing list.
• Execution tracking. Progress is tracked against the plan, so what slipped is visible instead of forgotten.
• On the tools you already use. It runs quietly in the background on the stack the business already has.

Falcor is a personal Chief of Staff agent, built on a reusable Agent OS framework — seven layers that turn a chat model into something that remembers, acts, and runs on its own.

• Seven layers. Identity and context define who it is and what it knows; skills define how it works; memory keeps what matters across sessions.
• Connected to the real world. It reaches email, calendar, documents, and chat through a connection layer — and a verification layer checks its work so it doesn't quietly get things wrong.
• Runs unattended. An automation layer handles scheduled jobs — briefings, check-ins, reports — without being asked.
• Reusable by design. The same framework can stand up a new agent for a different role or business.