Read the page, not just OCR
Charts and diagrams are passed as rendered page images so the model reasons over the visual source.
Vision-native compliance intelligence for payments operations
A citation-grounded document copilot for payments compliance, operations, and risk teams. It answers across regulations, SOPs, reports, and specifications—including charts and process diagrams—and links every claim to the exact source page and region.
Role · Product architecture and full-stack engineering
System anatomy
The problem
Payments teams work across dense regulations and internal documents where a plausible answer is not enough. They need visual understanding, exact citations, grounding verification, PII controls, and an auditable record of every interaction.
What the system does
Cross-document questions over regulatory, risk, SOP, and specification corpora
Vision-native reading of charts, tables, and process diagrams
Clickable citations that jump to and highlight the exact PDF region
Verifier-generated grounding score with low-confidence disclaimer
PII redaction before audit storage and tamper-evident interaction records
Compliance report export plus query, latency, confidence, and cost dashboards
Architecture
PyMuPDF page rendering and text extraction at ingest
BM25 + dense hybrid retrieval over page summaries and embeddings
Claude vision answer pass followed by a separate grounding verifier
ChromaDB for vectors, SQLite for metadata and audit, page-image cache on disk
Next.js workspace with react-pdf overlays, Zustand, TanStack Query, and SSE
Key decisions
Charts and diagrams are passed as rendered page images so the model reasons over the visual source.
Document, page, bounding box, quote, and reasoning are structured objects that drive navigation and audit.
Synthetic documents and explicit production-readiness gaps keep the demonstration honest and reviewable.