Bank Statement Analysis PoC

Document AI for a national digital bank's evaluators, with strict access control, data at rest in an Indonesian Azure region and faster document processing.

~20% faster
Document processing: mean run 233.9 s to 186.8 s; slowest run -34%
748 tests
Automated tests supported by 34 invariant verifier scripts
8 evaluators
Bank evaluators representing 3 personas used the environment

Role: Lead engineer, responsible for maintenance
Period: Jun 2026 to Sep 2026
Client: National digital bank
Project type: Featured
Room in the 3D HQ: L3:bank-statement-analyzer

Results

~20% faster
Document processing: mean run 233.9 s to 186.8 s; slowest run -34%
Replay benchmark over 4 runs with data stores wiped clean; all 4 retained 99.3 to 99.4% analysis confidence, with accuracy unchanged.
748 tests
Automated tests supported by 34 invariant verifier scripts
Enforced during builds and checked with mutation testing.
8 evaluators
Bank evaluators representing 3 personas used the environment
Used during hands-on PoC evaluation; the outcome is not public.

Problem

A national digital bank's evaluators needed auditable bank-statement analysis within a fixed PoC window. Uploads and results required strict access control, with data at rest kept in Indonesia.

Contribution

I built the platform and infrastructure as code in a fixed Indonesian Azure region.
I hardened sign-in with Entra OIDC, PKCE, JWKS verification and fail-closed role mapping, then added maker-checker controls, audit trails and three-language support.
I fixed critical defects before evaluator access: an SSO role-mapping flaw, an inverted CSRF check behind the platform proxy and demo paths capable of fabricating analysis.
I enforced team rules through build-time verifier scripts, bounded the document-processing page pool and contributed to the PoC technical proposal.