Gov Assistant

Released Production

A GenAI-powered tool that lives on the GenAI portal, focused on governing documents (Policies and Instructions). Every answer comes with citations that point straight to the relevant section of the source so compliance teams can act quickly and confidently, without second-guessing where the answer came from.

Python RAG Azure OpenAI LangChain Vector DB FastAPI AWS

Key Highlights

  • Live with the Risk & Compliance Tribe, serving Policies and Instructions and regulatory data and DoIT business procedures from internal and external documents.
  • Cut document ingestion time by 75% with an async, parallel embedding pipeline.
  • Built a citation engine from scratch that traces every answer back to its exact source section.

Gov Co-Author

Active Development Hot

A GenAI-powered Gov Co-Author that drafts, simplifies, and merges governing instructions. Where overlapping instructions need to come down to make it coherent. Powered by llm with mode Ask, Agent, and Canvas using GPT-5.1 and a custom generation pipeline that puts out a 200-page document in under 2 minutes.

Python GPT-5.1 LangChain Azure OpenAI FastAPI Canvas

Key Highlights

  • Moved to GPT-5.1 cut 200-page generation from 20 minutes to under 2.
  • Sections write in parallel, with template caching to avoid repeating work.
  • Spots overlaps and gaps against the FCC Policy framework on its own.
  • PoC delivered to Group Compliance; full build is in active development.

RAG Evaluation & Observability Framework

Completed

An internal harness for measuring how good our RAG pipelines really are retrieval quality, answer faithfulness, and citation accuracy across the board. Plugs into Kibana for live production monitoring, with alerts when answer quality starts slipping.

Python RAGAS Kibana Elasticsearch FastAPI

Key Highlights

  • Automated scoring for faithfulness, context recall, and answer relevance.
  • Live Kibana dashboards tracking retrieval and generation, query by query.
  • Made tuning chunk size, overlap, and re-ranking a numbers-driven decision instead of a guess.
OFF
ON