BlackSwan · Banking ML

Banks that learn
without sharing.

A federated credit-risk cockpit for banking pilots. Each bank trains on its own shard of real Lending Club data; the center only ever sees scores, confidence, and explanations — and Calybris turns them into APPROVE, REVIEW, or DECLINE.

bank shards risk evidence decision
01 — THE PROBLEM

Banks can't pool their data — but they still need shared signal.

A single bank's default history is thin and biased toward its own customer base. Pooling raw borrower rows across banks would give a far better risk model — but regulation, privacy, and competition make that impossible. So each institution is stuck learning credit risk alone, on a narrow slice of reality.

02 — HOW IT WORKS

Train local, share evidence, decide central.

01 · YEREL

Shards train alone

Income-segmented banks train local models on their own data — raw rows never leave the bank.

02 · KANIT

Only signal moves

The center consumes scores, confidence, drift, and explanation summaries — never borrower rows.

03 · KARAR

Calybris makes the call

The boundary turns PD, stress PD, confidence, and thresholds into deterministic APPROVE / REVIEW / DECLINE.

03 — WHAT IT SHOWS

Regulator-explainable, on real data.

Real data

Built on real Lending Club data, not synthetic telemetry.

Privacy by design

Raw borrower rows never cross a bank boundary — only evidence does.

Audit trail

Every decision bound to a Calybris audit boundary and stress-tested with fault injection.

Building something at a bank?

Let's talk about federated, explainable, audit-bound ML.

Start a conversation →