Crucible — the autonomous AI underwriter

Every lender's AI ends a hard loan with one word: no. Crucible is an autonomous AI underwriter that works the file until it's a yes. A loan is pulled straight from the system of record (Salesforce) and put on trial before three adversarial agents: a Prosecutor that files every reason to deny — each cited to retrieved guideline text (Fannie Mae, HUD 4000.1) — a Defender that answers every charge with a guideline-valid path, and an Adjudicator that rules on the cited record and issues a ranked "Path to Yes" with Reg-B adverse-action reasons for anything still blocking. The verdict is where Crucible starts, not stops. It writes the underwriter's real conditions report with industry codes, reads incoming documents and auto-clears the conditions they satisfy — with sufficiency math, so a $900 bank statement cannot clear a $31,400 funds condition — plans outreach to title, insurance, and borrower over Twilio rails, and submits down the Fannie/Freddie AUS seam. When a file is denied, its restructure solver searches buydowns, product switches, and down-payment moves to rebuild the deal. It learns from every file: approval scoring, restructure success rates, and which documents clear which conditions. AMD is load-bearing, not decorative: guideline RAG, deterministic underwriting math, adversary scans, and document parsing run locally on an AMD Instinct MI300X (vLLM on ROCm) — borrower PII never leaves the box — while only the hardest reasoning escalates to Fireworks. A live router ledger shows ~60% of calls handled on-device and the cost vs. an all-frontier pipeline. The console ships with a voiced guided tour: Crucible himself walks the platform, pulls a live file, and underwrites it in front of you — in seconds, not days.