Company Brain — Enterprise RAG Platform

Institutional knowledge in a lending business is scattered — across documents, policy notes, operational runbooks, ticket history and the heads of a few long-tenured people. Answering a question that spans two of those sources means knowing who to ask, and the answer is only as current as that person's memory. As teams grow, that pattern stops scaling and starts producing inconsistent decisions.

A knowledge platform built on retrieval-augmented generation. Sources are ingested and normalised into a consistent representation, retrieval combines semantic and keyword matching so exact identifiers still resolve, and answers are generated strictly from retrieved context with citations back to the source. Where retrieval confidence is low, the system declines rather than guessing.

• Architected the platform end to end and took it to production.
• Designed the ingestion layer that normalises multiple source formats into one retrievable representation.
• Built the hybrid retrieval strategy so semantic search and exact-identifier matching both work.
• Implemented citation-backed answering, with an explicit decline path when confidence is low.
• Established the evaluation loop that catches retrieval quality degrading as the knowledge base grows.

• Multi-source knowledge ingestion and normalisation
• Hybrid semantic + keyword retrieval
• Citation-backed answers traceable to their source
• Explicit low-confidence decline instead of plausible guessing
• Access-aware retrieval filtering
• Continuous evaluation against a held-out question set

• Python
• Django
• RAG
• LLMs
• Vector Retrieval
• AWS

• Centralized multi-source knowledge retrieval across functions
• Cross-functional decision support grounded in citable sources
• Reduced dependence on individuals as single points of knowledge