Diagnostic
two weeks, fixed feeEach produces a graded findings document and a value case in your own numbers — a deliverable on its own terms, not a sales exercise. Whatever follows is scoped and quoted inside it.
We build evidence systems and AI that survives being checked — for supply chains, trade operations and regulated workflows. Not generic AI consulting. The specific thing we have already built four times, in production, for our own products.
Each produces a graded findings document and a value case in your own numbers — a deliverable on its own terms, not a sales exercise. Whatever follows is scoped and quoted inside it.
Five costed packages. None is bespoke. Acceptance criteria are agreed before the SOW is signed, and acceptance is objective — a case set and pass thresholds both sides saw in advance.
Work that continues after a build lands, or that runs on its own cadence.
Diagnostic first. Always. We do not quote a build before we have looked at your data. If there is nothing worth building, both sides find out in two weeks rather than four months.
Two lenses from a single document pull. The money lens finds duty overpaid or never claimed, and leakage in accounts payable. The evidence lens grades how well your origin record would hold up if somebody asked for proof.
If your accounts payable and your broker come back clean, the report says so and the fee is refunded.
One live AI path, reviewed against a defined auditability standard and graded finding by finding. Can it cite at span level? Does it abstain when retrieval fails? Is grading independent, or is the model marking its own work?
We also state the limit that cannot be engineered away: for any open-corpus retrieval path there is a clause that was never retrieved.
Supplier master, purchase order and BOM ingest. A screening connector over published lists plus a source you license. A bilingual supplier evidence portal, four-grade provenance on every claim, a signed and hash-chained record, and a PDF dossier with an evidence index.
Not included: our own entity-risk database, legal representation or customs filings, any outcome guarantee.
Retrieval and grounding over your existing systems. Cell-level traceability. Signed answers carrying the model-version record. Explicit refusal behaviour. An evaluation set and regression harness handed to you. Role-based entitlement.
The broadest package here — neither apparel-specific nor trade-specific.
Passport data flowing into the systems you already run, built to the CEN standards cited in the Official Journal in July 2026. Mapping from supplier-side records to your schema, a conformance test suite, and a connector maintained against standard revisions.
For a brand already running a traceability platform, this is how supplier evidence reaches it without replacing anything.
The canonical record underneath everything else: entity resolution across systems, a claim model with evidence grades, lineage that survives a migration, and the write-through discipline that stops a second database drifting away from the first.
The least glamorous of the five. It is also the one most often skipped before an AI programme fails.
A signed record of what an agent did and under whose authority. Entitlement-aware tool access, action-level audit trail, scoped and revocable authority, and a replay capability so a decision can be reconstructed months later.
As systems start taking actions rather than answering questions, that record is the whole audit.
Your suppliers answer the same questions in a different form for every buyer. This maps one canonical factory record onto the frameworks you require, pre-fills the dossier, flags gaps as gaps, and leaves a person to review and approve.
Coverage is stated as built, never as intended. Two frameworks tested end to end, one modelled and in validation, three in build.
Monitoring, list-version refreshes, evaluation re-runs, and model or policy changes as they land. The evaluation harness we hand over only stays useful if somebody re-runs it when the model underneath changes.
Offered at handover, never a condition of it.
Fifty percent on signing, forty on acceptance-test pass, ten at handover. No part of the fee is tied to your production date.
Objective. The case set and the pass thresholds are agreed during the diagnostic and written into the SOW before it is signed.
Engagements are led from our San Jose office with our international engineering team behind them.
We do not train models on it. Where screening intelligence is needed, you license the source and we build a vendor-agnostic connector — we do not resell a risk database.
We are not going to show you a wall of logos. What we will show you is the machinery: the grounding and abstention layer, the signing, the evaluation harness and the graded-evidence model are running code with a test suite, built for our own products before they were ever offered to a client.
We will walk you through it on a screen share, in the repository, before you spend anything. A case study tells you a thing worked for someone else. A walkthrough shows you the thing. Or start with the work: see where our systems have run
If what you need is on that list, we will say so in the first conversation rather than the fourth.
Fixed fee, a deliverable either way, and no build quoted until we have looked.