Electrical engineer by training, product leader by trade, repeat founder by habit.
The question I keep coming back to lately: how an agent gets authority to act on someone's behalf, and how that authority stays scoped, logged, and revocable. Most agent demos end with an API key and hope. That's the gap APOA is trying to close, as an open-source layer on top of MCP.
Kyle Kingsbury joked an AI agent might get talked into a $950 grandmother burrito by El Farolito's chatbot. I think he was being optimistic. So I built a working prototype where the boundaries live in a signed token between the LLM and the execution layer, not in the prompt.
I gave an AI agent access to act on my behalf a few months ago. Within ten minutes I was scared of it. Not because it did anything wrong. Because I had no way to guarantee it wouldn't. The entire trust model was a system prompt and vibes. I built APOA as my take on a solution.
Always liked UI/UX design, and love working in Figma. Tried Claude Design recently and was so impressed with the quality of the output: multiple design directions, high-definition mockups. Makes me wonder how design will evolve over the next 12 months.
The piece I keep adding to MCP is capability attenuation: every time an agent delegates to another agent, the new permissions can only be narrower than the parent's. Never wider. The protocol enforces it at signing, and cascade revocation means pulling one token breaks the chain below it. Around that: OAuth for APIs, browser-based credential injection for the long tail that doesn't have one, natural-language soft rules logged next to hard rules the protocol refuses outright, per-action audit. None of it is novel alone. The question is whether it's the right default for agents acting on behalf of people.
Open-source authorization for AI agents, modeled on power of attorney. Works for services with APIs (OAuth under the hood) and services without (browser-based credential injection where the AI never sees the credentials). TypeScript and Python SDKs shipped, plus MCP and A2A packages.
An open protocol where the services an agent calls sign encrypted receipts for what they observed, then publish them to an append-only transparency log. The owner verifies later with their own key, so a compromised agent can't rewrite its own trail. Built on COSE signatures, HPKE encryption, and SCITT-style logs. The mental model: Find My, but for AI agents.
A pre-flight manual for shipping AI-assisted software. A five-level severity model (from a weekend toy to a hospital intake form), a 25-point pre-flight checklist, thirty responsible prompt templates, and seven postmortems from recent incidents. Five minutes, five questions, every deploy. Written for first-time builders and senior engineers reviewing their agent's hundredth PR alike.
A fun take on the iconic Street Fighter game I grew up with, starring tech CEOs instead of Ryu and Ken. Built over Thanksgiving weekend with Claude Code, Cursor, and Nano Banana Pro.
Leading AI personalization and membership products inside the Amex mobile app. Built and mentored a team of product managers across membership and AI initiatives.
B2B agricultural fleet management platform. Built and scaled from zero to acquisition over ten years. Most of what I know about running a real business came from here.
Built the fitness recommendation engine that increased user engagement by 25%. Optimized backend API performance by 20%.
Architected incident management platform for 24/7 monitoring of high-frequency trading systems. Wrote business requirements and UX flows for financial products.
Coordinated transition of 5 RF semiconductor products from experimental stage to mass production, resulting in a $4M order for an advanced semiconductor developer.
Always looking to meet interesting people!