01 / Domain
AI × Crypto
AI you can verify. Agents you can trust with keys.
The problem
Agents are about to move real money. Without proof of what ran, and for whom, AI on-chain is just a new attack surface.
The next wave of on-chain activity won't be typed by people. It will be proposed, negotiated and signed by software. That shift is useful only if every action can be explained after the fact: which model, which policy, which permission, which human said yes.
We design AI systems the way we design payment rails. Bounded authority, verifiable execution and an audit trail you can hand to a regulator, a counterparty or your own board.
Where it applies
- Treasury copilotsAgents that rebalance, hedge or sweep funds inside hard limits, with escalation for anything unusual.
- Trust-minimised oraclesModel outputs delivered on-chain with attestations a contract can verify before acting.
- Support and operations agentsAI that triages incidents and drafts fixes, while engineers review and merge.
- Data and model marketplacesLicensing and provenance for datasets and weights, enforced in code.
What we build
Verifiable inference
zkML and TEE pipelines that prove which model produced an output.
Agentic wallets
Policy-bound agents with spend limits, session keys and human approval.
Model provenance
On-chain registries for weights, datasets and licences.
AI-assisted ops
Copilots for infra and incidents, with humans owning the merge.
How we approach it
- 01Threat-model the agent before the prompt
- 02Every action signed, scoped and logged
- 03Evaluate offline, then shadow in production
Stack
What you get
- ✦Agents with explicit, reviewable authority
- ✦Evidence of what ran, for every action
- ✦Faster operations without giving up control
Have a hard problem?
Tell us what you are building. We reply within two working days.