A senior credit memo used to take an analyst five days of spreadsheet archaeology. This LLM copilot drafts a cited, decision-ready memo in four hours -- and shows its work on every claim.
Underwriting can't run on a model that sounds confident and cites nothing. Every quantitative sentence in the memo is anchored to a pgvector source span -- the exact financial line or KYB filing it came from -- so reviewers click a claim and land on its evidence.
A verifier node closes the loop: it re-reads each cited chunk and rejects any sentence the source doesn't support, sending it back through the LangGraph state machine for a grounded rewrite rather than shipping a plausible guess.
The copilot pairs a gradient-boosted probability-of-default model (AUC 0.91) with the LLM's narrative read. The model supplies a calibrated, defensible grade; the LLM must reconcile its prose against that number rather than freelancing a verdict.
When the narrative and the score disagree, that tension surfaces to the analyst as a flag -- turning a black-box output into a structured prompt for human judgment exactly where it matters most.
| Stage | Before | After |
|---|---|---|
| Data gathering | Manual spreadsheet pulls, 1-2 days | FastAPI ingest + pgvector, minutes |
| Risk scoring | Analyst rules-of-thumb | Calibrated PD model, AUC 0.91 |
| First draft | Written from scratch | LLM drafts 75% of sections |
| Citations | Hand-traced footnotes | Auto-anchored to source spans |
| Verification | Peer re-reads full memo | Verifier node flags unsupported claims |
| Turnaround | ~5 business days |
Every figure and covenant claim links to its source span in the underlying financials or KYB filing.
A gradient-boosted default model (AUC 0.91) anchors the narrative to a defensible numeric risk grade.
Section drafting, retrieval and verification run as explicit graph nodes, not an opaque single prompt.
A self-critique node re-checks each citation and regenerates any sentence the source doesn't support.
Source-anchored embeddings let retrieval pull the precise ratios and red-flags each section needs.
Statements, bank feeds and KYB documents are normalized and indexed through a single typed API surface.
“It doesn't write a memo that sounds right -- it writes one where every claim points back to the line it came from.”
The AI Underwriting Copilot turns a five-day analyst slog into a four-hour, evidence-backed draft -- automating 75% of memo sections and cutting manual effort by 70% without surrendering auditability. By pairing a calibrated 0.91-AUC default model with a citation-verified LLM drafting loop over Python, Claude, LangGraph and pgvector, it keeps humans in the seat for the judgment calls and out of the spreadsheet archaeology. The result is faster credit decisions that a regulator -- or a skeptical credit committee -- can actually trace.
| ~4 hours |