Falcon

VERITAS — The AI Forensic Auditor. Reads 100% of a company's books, files a cited fraud examination in about a minute, for a cent. Companies lose 5% of revenue to fraud (~$5T/year, ACFE). External audits catch just 3% of it — because humans sample: they read ~1% of the books and hope. The most common fraud on earth is painfully simple — an employee registers a shell vendor at their own home address, self-approves its invoices, and drains the company. The proof is in the books, scattered across documents nobody reads side by side. VERITAS reads all of them. Attach a folder of a company's books — 1,090 real documents in our demo: invoices, bank statements, payroll registers, HR records, vendor registrations, board minutes, credit notes — and ask it to audit your company. In about a minute it files a forensic examination: €635,400 at risk across three fraud schemes (a shell company registered at a procurement manager's home address, a ghost employee paid into a colleague's bank account, a duplicate payment with no reversing credit note) — with every claim cited to a clickable source document, and a downloadable examination report. Total inference cost: ~$0.01. Why it's a genuine agent, not a pipeline. For every lead, on screen: VultronRetriever Core surfaces the evidence → the examiner streams its fraud hypothesis AND the innocent explanation it must rule out → it writes its own follow-up retrieval query, which goes to VultronRetriever Prime → it streams a verdict weighing both retrieval rounds. Plans once, retrieves twice per lead, decides every time. The filing rule (our favorite part): no accusation on one model family's say-so. A finding is filed only when the Qwen examiner confirms it AND an independent panel of three NVIDIA Nemotron reviewers — a different model family, with no fallback to the examiner's model, so their blind spots can't correlate — upholds it across three adversarial lenses: correctness, the innocent explanation, evidentiary sufficiency. A refused accusation is not filed. If the examiner contradicts dispositive document evidence, the panel arbitrates. Reviewers that error abstain — and an abstention never upholds an accusation. It can't cry wolf. A fourth suspicious duplicate in the demo books was investigated and cleared — a credit note had reversed it. Our live eval suite runs the full pipeline on clean books with the fraud proofs surgically removed: it files zero findings. Dollar amounts are never generated — every figure is recomputed deterministically from the ledger. There's a second bundled company (Kestrel Manufacturing, USD) proving the same agent works on books it has never seen, in another currency. Built 100% on Vultr. VultronRetriever Prime + Core for retrieval (/v1/rerank), Qwen 3.6 as the examiner, NVIDIA Nemotron Cascade-2 as the drone extraction fleet + the binding review panel — all on Vultr Serverless Inference — served from one Vultr Cloud Compute VM. Refresh-proof SSE streaming, spend kill-switch, rate limiting, and a live interrogation mode: ask "how do you know it's a shell?" and it re-retrieves from the actual corpus and answers with citations, streaming. Try it live: https://veritas.144-202-6-174.sslip.io — upload your own books; on clean books it files nothing. Every line written during the event.