The challenge
A leading global financial institution had banking knowledge, policies, adviser notes, compliance procedures and meeting transcripts, fragmented across legacy intranets, SharePoint, shared drives and team wikis, each with different access rules and update schedules. Staff spent five to seven hours a week searching multiple systems for authoritative answers, keyword search returned unranked document lists without synthesis, and inconsistent answers had been flagged by internal audit as a regulatory control weakness.
What we did
Delivered a governed multi-agent RAG platform on Databricks: a Delta medallion architecture ingesting documents, notes and transcripts, document parsing and named-entity recognition, semantic chunking, and Vector Search with Delta Sync indexing. LangGraph orchestration classifies each query by intent and routes it to specialised retrieval agents; fine-tuned financial embeddings improve recall for domain terminology, and the synthesis layer adapts prompting strategy by query type, extractive for policies, abstractive for summaries. Every response carries source links, versions and a confidence score, evaluated continuously with DeepEval Faithfulness metrics in MLflow Evaluate against compliance-curated golden datasets, with subthreshold responses flagged for human review, all behind AI Gateway controls and a foundational authentication model.
The outcome
Document retrieval that previously took hours now returns sourced, versioned answers in under four seconds. The answer inconsistencies that regulators had flagged as a control weakness were eliminated, and the architecture became reusable across multiple banking domains including treasury, trade finance and customer onboarding.
Built a governed document-to-AI foundation that turns fragmented banking knowledge into sourced, evaluated answers in under four seconds.
Unstructured content landed in a medallion lakehouse, enriched via NER and chunking, indexed for Vector Search, and served through LangGraph-orchestrated agents behind an AI Gateway. DeepEval Faithfulness metrics run continuously against compliance-curated golden datasets in MLflow Evaluate, flagging subthreshold responses for human review before they reach a banking user.