An AI company
for onchain finance

Plenty of complicated systems have usable interfaces. What makes onchain finance different is that the consequences of a small mistake are total and immediate. A mistyped address is not a validation error; it is a permanent loss. An unlimited approval is not a preference; it is a standing claim on your balance that outlives the transaction that created it.

So the interface problem is really a verification problem. Users do not need fewer fields. They need something that checks the fields for them, and can explain what it checked.

A language model reads intent very well. Given “move two thousand dollars of USDC from Base to Solana, safest route”, it will correctly identify the action, the asset, the networks and the priority. That is a genuine advance over a form.

What a model cannot be is the thing that decides an address is correct. Or that a token has six decimals rather than eighteen. Or that a balance covers an amount. Those questions have exact answers, and a system that answers them probabilistically will be right almost every time — which is precisely the failure mode that costs people money, because almost every time is indistinguishable from always until it is not.

ZeFi splits the work along the line where the two kinds of question actually divide. Interpretation is a language problem and gets a language model. Verification is an engineering problem and gets deterministic code: a curated registry, checksum validation, BigInt arithmetic, an explicit status machine.

Between them sits an artefact — the transaction plan — that a person can read. It states what ZeFi understood, what it assumed, what it will do, what it will cost, what it cannot do, and where every number came from. The plan is the product. The chat is how you get one.

A team deciding which chain to expand to is doing route selection with worse information and higher stakes than an individual bridging stablecoins. The same discipline applies: build a structural model, score the options against it, name the risks, and produce something a person can act on.

ZeFi collapses many possible routes into one verified action. Routefold takes one product and unfolds it into ranked destinations. Same company, same method, opposite direction of travel.

Eventually a user should be able to set a policy — never more than a thousand a day, only these contracts, never an unlimited approval, always keep gas — and let bounded automation work inside it. That is a real product, and it is not this one yet.

It depends on smart accounts, session keys and deep simulation, because a policy is only meaningful if something other than the agent enforces it. Building the execution layer first and the constraints afterwards is how you get an autonomous system nobody should use. We are doing it the other way round.