A compliance gate fused to a supply chain
A subvendor marketplace for a staffing firm, live in production: AI document review feeding a deterministic rule engine, humans on the exceptions, and a placement lifecycle riding an append-only event spine behind it.
- email to decision, measured end to end
- requests, zero failures, 200 VU
- test cases gating CI
The problem
A staffing firm’s compliance desk lived in an inbox. Certificates of insurance, work authorizations, subvendor agreements: every document read by a person, checked against rules held in someone’s head, chased across email threads. Slow when it worked; expensive when it didn’t, because a missed document is a compliance exposure with a company’s name on it.
The system
The pipeline inverts the human’s job from reading everything to reviewing exceptions.
Inbound documents are classified and their fields extracted by an AI review layer. The decisions are not made by the model. Extracted facts feed a deterministic rule engine that renders the compliance verdict the same way every time; the AI reads, the rules decide. Anything ambiguous or below confidence lands in a human-review queue with the evidence attached.
Approved vendors flow to e-signature and sync into the staffing ERP. Commercial terms are enforced in code, at the platform layer, not in a spreadsheet someone remembers to check.
The measured result: an email arriving to a compliance decision in about five seconds, end to end, on the production deployment.
The marketplace behind the gate
Compliance is the door; the supply chain is the building. Open requisitions release only to partners who have cleared the gate. Submitted candidates run an advisory AI pre-screen into a recruiter queue, and every accepted candidate rides a placement lifecycle, from interview through start to termination, on an append-only event spine. A hard ownership lock ends cross-vendor double-submits: one firm per candidate per client, enforced with a constraint, not a policy memo.
One number the platform never shows anyone: a partner’s bill rate stays out of the portal and out of every model prompt. The margin firewall is structural. The AI advises on documents and candidates; it never touches the money.
Ahead of go-live the platform held a staged load baseline: 106,000 requests with zero failures at 200 concurrent virtual users, and 1,160 test cases gate its CI on every push.
Receipts
- Email to decision in ~5 seconds, measured end to end on the production deployment (2026-07)
- 106,000 requests, zero failures across the staged 200-VU load baseline (2026-07)
- 1,160 test cases gating CI: server and UI suites, both wired into the pipeline (2026-07)
What this means for you
Somewhere in your operation a person is reading documents a system should be reading, and making calls a rule engine should be making identically every time. We build exactly that flow: AI where judgment reads, deterministic code where policy decides, humans on the exceptions. Fixed scope, and if it isn’t deployed and working, you don’t pay the back half.