01Agentic AI · RAG · Compliance
Banking Compliance Agent.
An agentic AI system for reviewing lending documents against financial regulation, surfacing evidence, uncertainty and cases requiring human review.
01 · Problem
Compliance review is slow, manual and evidence-heavy.
Loan and policy documents must be checked against a large regulatory corpus. The challenge is not simply generating an answer; it is retrieving the right evidence, exposing uncertainty and making escalation explicit.
02 · Approach
RAG first, tools second, human judgement preserved.
The system combines retrieval over FCA/PRA material with tool access for policy and customer context, a structured compliance reasoning layer, a deterministic confidence gate and an auditable report.
03 · Decisions
Designed for trade-offs, not demos.
RAG was preferred to fine-tuning for updateability and traceable evidence. A single orchestrating flow with validation was favoured over unnecessary multi-agent complexity. Tool access is deliberately bounded.
04 · Evidence
Evaluation goes beyond fluent answers.
Retrieval quality, grounding, abstention behaviour, latency and rule-based checks are treated as first-class evaluation targets. The system is designed to fail visibly when evidence quality is insufficient.
05 · Reliability
Trust is part of the architecture.
Human review is triggered by uncertainty. Outputs preserve source traceability, deterministic checks and an audit trail so a confident-sounding answer is never treated as sufficient evidence by itself.
06 · Reflection
The hardest failure was silent, not dramatic.
An ingestion pipeline initially produced thousands of chunks while preserving raw PDF content. The lesson was to validate content quality, not just pipeline completion—successful execution does not imply successful output.
System architecture