An open protocol for what an AI agent is allowed to do, and proof of what it did.
It checks authority before an action runs and leaves a signed record after. The protocol, the gateway and the SDKs are open source, and the spec is an IETF Internet-Draft.
My attempt at artificial creative intelligence, ACI.
An AI that creates and publishes for you. You talk, the system turns it into content across LinkedIn, Instagram, TikTok and YouTube, and agents handle the rest. We reached 300+ users and 1.6M monthly impressions.
A sustainable straw built for industrial scale.
I started ofNature after Starbucks told us our first sustainable straw could not scale. We built a cellulose acetate alternative, raised $150K and set up manufacturing in Ukraine. The war ended the company. In 2025, McDonald's switched to cellulose acetate straws.
Things I build because I want to know what happens.
Meeting people through AI instead of feeds. Tell your agent who you want to meet. It finds matches, and both people opt in before anything happens.
ExploreIf AI companies make money from your data, should some of that value come back to you? You choose what to license. If it sells, you get paid.
Learn moreA nose wearable with airflow sensors and an AI coach, built to make breathing measurable. The hardware worked. I paused it when the next step needed a medical cofounder.
Full project67 agents monitoring one patient at once. The question was whether agents that update each other's beliefs catch risks that a simple average misses.
Watch demoSmall apps and weird ideas I build fast to see if they deserve more time. Some become products. Some are a weekend well spent.
ExploreI use AI to make music, because some stories are easier to sing than explain. Ten albums in two months, a lot of them about love that never happened.
ListenBefore startups, I directed films and music videos. Some reached 14M+ views. I still come back to visual storytelling when code is the wrong medium.
View allPublished papers, drafts, and ongoing work on AI agent governance.
Defines the three-layer protocol — Agent Passport, Human Values Floor, Beneficiary Attribution — that grounds the Agent Passport System.
Authority attenuation formalized — delegation can only narrow, never expand. Formal invariants and adversarial testing.
Formal model for how agent authority narrows monotonically across delegation chains.
Lifecycle controls for derivative AI artifacts — telemetry scopes, behavioral measurement objects, bring-your-own-model boundaries.
Binding physical-layer attestation to delegation chains. Seven experiments on IBM Quantum showing measurable Bell and GHZ fidelity gaps.
The unit of agent governance is the population-with-medium, not the individual agent. Six rounds of adversarial review.
Cryptographic commitment over which sparse-autoencoder features were active during a decision. Envelope spec, three-stage verification, four dispute primitives. Experimental results against Llama-3.1-8B via Neuronpedia.
Cryptographic governance proves procedural validity, not effect safety. Five omitted-variable classes catalogued, defeat traces in reference implementation.
A controlled isolation of the retention rule in populations of short-lived LLM agents. Peer-voted retention kept a worse shared memory than random selection in 4 of 5 replicates. Pre-registrations, kill conditions, and artifact record included.
IETF Internet-Draft formalizing the Agent Passport System interoperability surface for cross-protocol agent identity, delegation, and enforcement.