Which version of the rule applies in our state? Did it change last month? Where's the source?

A RegTech team needed their users to get those answers without searching federal, state and local databases by hand. The rules were public. Finding the one that applied to a specific jurisdiction, and knowing it was still current, was the slow part.

So we built the system underneath the question box:

• Automated ingestion of web pages and PDFs from 1,994 federal, state and local sources, eCFR included • A jurisdiction-aware knowledge base on PostgreSQL and pgvector, using local MiniLM embeddings • Answers generated with Claude and grounded in the retrieved rules, with a citation on every claim • Scheduled change detection that re-checks sources and re-ingests the ones that changed • Tiered entitlements, Stripe billing, analytics and audit logging, running on Azure Container Apps with Service Bus

The result: 35,643 rules from 1,994 unique sources in one searchable, monitored knowledge base. It answers in plain language and shows where every answer came from, and it's live in production.

Most of the effort went into ingestion and monitoring, not the chat interface. An answer is only useful if the rule behind it is current and scoped to the right place.

Building AI on regulated data? Message me and we can talk it through.

#RAG #RegTech #AIEngineering #LegalTech