~40s
One command, from cold start to a sealed case file with a real proof inside
Tamper-evident AI audit trails, powered by Memtara zero-knowledge proofs. Real Baby Jubjub signatures. SHA-256 hash chains. Built against DFSA COB 3.4.
That gap is where AI governance in regulated finance actually fails. A prompt asserting "this client qualifies for the note" gets chained faithfully, signed, and stored — and it is still prompt text, which is attacker-controlled. AIHOOTS is the governance platform; Memtara is the zero-knowledge engine that closes the gap underneath it.
On the client's own device, a zero-knowledge proof is generated over a vault the client already holds, against terms the bank registered in advance. Baby Jubjub signatures, Poseidon commitments, a Noir circuit, a Barretenberg proof. The figures are the witness — they stay on the device by construction, not by policy.
Memtara verifies the proof against a committed verification key and issues an Ed25519-signed token that lives for 300 seconds. The registry rejects terms supplied by the caller, so a bank cannot quietly move its own goalposts.
The audit gateway checks the attestation against a published JWKS with no synchronous call back to Memtara, then puts the verified fact into the model's context. The model is no longer being asked to take the prompt's word for it.
The proof hash is appended to a SHA-256 hash chain alongside the request, the response and the policy decision. An independent verifier detects a single flipped byte at any position — including one the platform itself flipped.
Each tab is a circuit that exists in the repository and compiles today. The status chip says how far it has been taken, because "we could build that" and "that is built" are different sentences and only one of them belongs on a website.
A bank must satisfy itself a structured product is suitable before recommending it. Conventionally that means the client hands over income, portfolio and risk-profile documents and the bank stores them — a permanent, breachable copy of a client's whole financial position, held to justify one recommendation.
This circuit produces the same regulatory evidence without the copy. The device proves all four DFSA COB 3.1 limbs were evaluated against the product's registered terms, and discloses exactly one bit: suitable, or not.
Proves years of experience meet a minimum and that a fixed number of achievement records genuinely exist in the vault, without revealing the record set. The pattern behind DIFC and ADGM onboarding journeys where a counterparty needs a threshold answered, not a dossier delivered.
Time-bound like every session circuit: a disclosure that cannot expire is a disclosure you have permanently made.
Proves a real income value falls inside a range and that a fixed number of deduction records genuinely exist, without revealing the exact figure or the underlying documents. The primitive behind proof-of-funds, accredited-investor certification and mortgage pre-qualification.
The counterparty learns that you clear the bar. It does not learn by how much, which is the part that ends up in a data breach.
The circuit written for exactly this platform. It proves that every record disclosed to a model is genuinely in the user's vault and falls under a category the user consented to share — turning "the AI never saw more than it was granted" from an assertion in an architecture document into a checked constraint.
Paired with the gateway's hash chain, the result is an audit trail where both halves are verifiable: what the model was told, and that it was true.
All use cases secured by the Memtara ZK engine.
A fifth circuit, emergency_session, covers time-boxed medical disclosure.
Crypto/VARA, card issuance and trade finance are frequently asked for and are
not built — they are roadmap, and this
page will say "live" on the day they are, not before.
Everything above is checkable in an afternoon by someone who does not trust us, which is the only kind of claim worth putting on a website with zero customers behind it.
$ git clone --recurse-submodules https://github.com/prasantk8/memtara-zkp.git $ cd memtara-zkp && ./scripts/quickstart.sh [1/6] registering the product and its suitability terms terms_hash 9f2c…a41d (registry rejects caller-supplied terms) [2/6] client device builds the witness figures never leave this step [3/6] generating the proof bb prove [4/6] verifying against the committed vkey VALID [5/6] issuing the attestation Ed25519, ttl 300s [6/6] sealing the Canonical Case File ✔ Canonical_Case_File.pdf 51 KB ✔ seal verified — file matches the digest recorded at export $ ./scripts/memtara-export verify Canonical_Case_File.pdf ; echo $? 0 # flip a single byte, then verify again 1
The AIHOOTS E1 audit gateway is an OpenAI-compatible proxy: point any OpenAI client
at it and every request, response and policy decision lands in a tamper-evident
chain, with an independent verifier that detects tampering at any position. It runs
on your infrastructure — docker compose up, then point your client at
localhost:8000/v1. There is no hosted proxy holding your prompts,
which for a regulated institution is the point rather than a limitation.
Every item here is something a competitor would leave out. They are on the page because the readers this page is for check things, and finding one of these on your own is worse than reading it from us.