Agentic engineer · founder of TRU · NYC
I build autonomous systems — and now the operating system they run on. I went from shipping 20+ autonomous systems in eight weeks to building the OS that lets one person run a whole company.
Before this: Trust & Safety at Google, ESCP, and a creative studio in NYC.
🧠 SynthOS — the AI operating system for a company of one. A native macOS cockpit where agents are the runtime, not a chat box: chat, terminal, browser, desktop, and voice from one command center, over a knowledge vault you own, with every file kept on your Mac. Multi-agent, in parallel, delivering the finished artifact — not a transcript. Hire no one. Run everything. → trusynth.com
⚒️ FORGE — turn a private vault into a live product. Publish it as a live MCP server any agent can call, or a one-line CLI — scoped read-only by default, token-gated, and revocable in one click. The vault never leaves your Mac. → trusynth.com/forge
🔐 Local-first by principle — your keys, your data, your business. Approvals, audit logs, and spend limits as first-class features.
TypeScript · Python · Swift (native macOS) · Solidity · Three.js / WebGL / GLSL · Claude SDK · MCP · Bash
The throughline across every repo below: autonomous multi-agent systems wired to real execution rails — agents that trade on-chain in private, wake from a voice command and ship, live inside 3D worlds, or open funded production rooms with no human in the loop.
| Project | What it is | Proof |
|---|---|---|
| SynthOS | The AI OS for a company of one — native macOS, agents as the runtime | Live · trusynth.com |
| HACKERA | Multi-agent World-Cup scouting — scored 665 players across all 64 matches of the 2022 World Cup | Live demo · self-reported 81% backtest (13/16) |
| genie | Say "genie" into a video → an ephemeral Claude Code instance wakes with browser + shell rails, ships the wish, reports back | Voice → autonomous-deploy pipeline · Claude SDK + MCP |
| ghost | Privacy-preserving AI DAO treasury — a Claude agent trades a shielded pool on Monad while members deposit openly | ZK unshield→swap→reshield · on-chain GhostVault.sol guardrails |
| oracle-society | AI agents that live inside photorealistic 3D Gaussian-splat worlds and reshape them mid-conversation | Three.js + splats · Gemini Live · MiniMax · MediaPipe |
| oracle-world-explorer | Navigate AI-generated splat worlds with head tracking + hand gestures — no controller, no headset, just a webcam | MediaPipe Hands/FaceMesh · raycast collision |
| greenroom | An AI production studio inside an encrypted group chat — 3 agents open funded rooms in under 90s, no human in the loop | XMTP encrypted chat · gateway agents · QR onboarding |
🛠 Also: claude-starter-kit — a drop-in Claude Code setup (auto/bypass mode, context-% statusline, done-chime). Plus 150+ shipped agent-generated micro-deploys not listed here.
🏆 Serial hackathon shipper — six agentic products across NYC's 2026 circuit
Six working multi-agent builds across NYC's 2026 AI/crypto circuit — one per weekend. The spread of sponsors is the flex: Unlink × Monad · Betaworks/MischiefClaw · TRAE × MiniMax · World Labs · Claw Hack. Every entry shipped a running build with real execution rails — not a slide deck.
| # | Event | Build | The hook |
|---|---|---|---|
| 1 | Ship Private. Ship Fast. — Unlink × Monad (Feb 27 – Mar 1) | ghost | An AI agent trades a shielded pool secretly on Monad while members deposit openly — ZK unshield/swap/reshield with 3-layer rogue-AI guardrails. |
| 2 | MischiefClaw — Betaworks NYC (April) | genie | Voice-triggered autonomous agent: a wish spoken into a video wakes a Claude Code instance that browses, builds, deploys, and reports back. |
| 3 | TRAE × MiniMax NYC (Mar 30) | oracle-society | AI agents that live inside photorealistic 3D Gaussian-splat worlds and reshape them mid-conversation. |
| 4 | World Labs Open World (March) | oracle-world-explorer | Gesture + head-tracking navigation of AI-generated splat worlds — webcam only, no headset. |
| 5 | Claw Hack NYC (February) | greenroom | 3 autonomous agents open a funded content-production room inside encrypted chat in under 90 seconds. |
| 6 | NYC AI Hackathon (April) | HACKERA | Multi-agent World-Cup scouting — self-reported 81% backtest on breakout-transfer calls. |
All six are accurately described as "built at" these events — working shipped products, not claimed prize placements. HACKERA's 81% is its own self-reported backtest (13 of 16).
Most of the work ships in private repos — public engagement metrics don't reflect it. The numbers above are the real signal; the card is the real stack (filtered past the static-deploy noise).
