Agentic AI that vets customer document packs (salary certificates, utility bills, tenancy and vehicle papers) for forgery, tampering and cross-document fraud.
KYC: Document Verification
Interactive demo running entirely in the browser on a real snapshot of 4 completed verification cases (synthetic documents). The agentic pipeline, forensics and portal checks run live in production on a Node + SQLite backend with a vision-language model and portal connectors.
Problem
High-volume onboarding and credit teams vet hundreds of document packs a day by hand: checking formats, validating against government portals, matching names across documents, and spotting reused or tampered files. It is slow, inconsistent, and the subtle forgeries are exactly the ones a tired reviewer misses.
Approach
A workflow spine with a few genuinely agentic joints: documents are classified and field-extracted by a vision model, run through deterministic rule checks (format, dates, amount-in-words), forensic analysis (PDF metadata, image re-compression, QR), cross-document name matching and reuse/velocity fingerprinting, then agentic nodes adjudicate forgery, materiality and verification-path recovery. Every disposition is auditable, evidence-cited and confidence-scored, and low-confidence cases route to a human vetter.
Outcomes
