days → minutesto read and cross-check bank statements, invoices, and tax filings
100%of decisions made by people — the AI never approves or rejects anyone
0.85disparate-impact ratio — passes the fair-lending four-fifths rule
How it works
- The AI reads. It pulls the real numbers out of the paperwork — every
figure linked back to the exact document — and cross-checks them: do the bank deposits
support the declared income?
- A transparent score. The credit result comes from a simple, explainable
model that states its reasons — never a black box, never a chatbot.
- People decide. An analyst reviews the file; a credit committee approves,
declines, or sets conditions. To go against the model's band, they must record a reason.
- Fairness is checked, not assumed. A disparate-impact report verifies the
scoring doesn't discriminate — and it's a report anyone can read.
The line that defines the product
Bridge draws a hard line: language AI reads documents and drafts the memo;
it never sets the score and never makes the decision. That boundary isn't a
promise — it's built into the code, so the applicant's documents and any personal
characteristic physically can't reach the scoring model. Lending decisions affect lives and
are heavily regulated; this is the kind of AI system where the right answer to "can the model
decide?" is no.
Proof it works
Bridge runs on a fully synthetic set of loan applications — realistic
paperwork, no real people's data. In testing it reads the figures back with perfect accuracy,
catches inflated income declarations that the bank deposits don't support (without falsely
flagging honest applicants), and a 60-business portfolio passes fair-lending testing. It's
also the first high-risk system registered in
Guardrail's compliance inventory.
For technical readers
- Document extraction attaches source provenance to every figure; cross-document reconciliation flags declared revenue > 1.4× bank inflows.
- Scoring is an additive, monotonic scorecard with adverse-action reason codes — no ML black box. Its function signature takes only structured features, so text and the protected-group label cannot reach it (a unit test proves the score is identical across groups).
- The memo drafter (template or Claude) cites evidence and leaves the recommendation blank; the human decision gate requires an override reason to depart from the score band and generates the adverse-action notice from the reason codes.
- Two Checkpoint eval gates: extraction accuracy (10/10) and fraud detection (10/10) on labeled synthetic cases.
What it can't do yet
- Extraction reads a synthetic document format; production would add vision/LLM extraction of scanned PDFs behind the same interface.
- The scorecard bands are hand-calibrated, not yet trained on a real lending portfolio.
- It's decision-support — a person underwrites; the model never does.