Independent AI Researcher & Engineer - AI systems architecture, multi-agent orchestration, MCP infrastructure. Founder of Peak Summit Labs.
I build AI infrastructure and ship it as products. Research becomes tools. Tools become papers. Papers validate the tools. One flywheel.
-
Multi-Agent Orchestration - Cross-machine agent coordination across multiple Macs via Cloudflare Durable Objects. 80+ skills, 38 hooks, terminal-spawned agents with backchannel messaging. Not a framework - a production system I use daily.
-
MCP Infrastructure - 16+ MCP servers built. Authored BMCPS v3.0, an open standard for safe MCP tool interfaces (25 mandatory tools, 238 automated tests, MIT licensed). Built a meta-server that generates new MCP servers from the standard.
-
Semantic Memory - Persistent memory system using PostgreSQL + pgvector. 10,000+ observations, cross-device sync, semantic search across sessions. Published research paper with 98.3% precision on LongMemEvalS benchmark.
-
Open Source - Contributor to Rectangle (25K+ stars). Added 8ths/9ths/12ths/16ths window divisions and position cycling system. Merged PR #1720.
-
Shipped Products - Two macOS apps on the App Store through Peak Summit Labs. Local-first, one-time pricing, no telemetry. I build what teams quote 9 months for - solo, using the AI infrastructure I built.
| Project | What It Does |
|---|---|
| reminisce | Cognitive science-inspired memory architecture for AI agents - working/episodic/semantic tiers with salience scoring. Paper | Benchmark Data |
| longterm-memory-macos | Semantic memory for Claude - PostgreSQL + pgvector with browser extension and dashboard |
| evm-chains-mcp-server | 111 tools across 7 networks - MCP at scale for complex multi-system orchestration |
| ccxt-mcp-server | Unified MCP interface to 106+ financial data APIs with real-time arbitrage detection |
| academic-lectures | Lecture materials from teaching at Ivy League institutions and Oxford |
Guest lectures at Harvard (x2), MIT Sloan, Princeton, Cornell Tech, NYU, and Oxford. Keynote at Web3 Summit Amsterdam. Eight talks in six months on AI systems, agent orchestration, and MCP infrastructure.
Full lecture portfolio: academic-lectures
Published:
- Reminisce: A Cognitive Science-Inspired Memory Architecture for AI Agents (2026) - Three-tier cognitive memory with salience scoring, evaluated on LongMemEvalS across 3 Claude model tiers. 98.3% precision. Published on Hugging Face.
Interests:
- Cognitive science-inspired memory architectures for AI agents
- Multi-agent cooperation and task decomposition
- Tool-call runtimes and safe execution patterns
- Context engineering and prompt architecture
| Product | What It Is |
|---|---|
| Push to Transcribe | Voice-to-text for macOS. Hold a key, speak, release - text appears. Whisper-powered, fully local. $19.99 one-time. |
| MarkDossier | WYSIWYG markdown editor for macOS. Hot-reloads when agents write to your files. Token counting built in. $14.99 one-time. |
45+ posts at operationalsemantics.dev documenting the full journey from MCP experiments to published standard and multi-agent systems. Includes an 8-part series on the Ivy League lecture tour.
Languages: TypeScript, Python, Swift, Rust AI: Claude, MCP, Ollama, LM Studio Infrastructure: Cloudflare (Workers, D1, R2, Pages), PostgreSQL, pgvector, bun
Blog - BMCPS Standard - Portfolio - Peak Summit Labs - LinkedIn - X