← Previous case study
Case Study05 / 05
ClientIndependent — Open Source
RoleCreator, Product & Engineering
StackPython · Claude · Agentic pipelines
Status2026 — ongoing
Agentic AI · Responsible AI

Proving the controls-by-design pattern travels beyond one bank — by shipping it as code.

The brief

The governance pattern that secured CCO sign-off at Deutsche Bank — explainability, a full audit chain, human-in-the-loop review, confidence gating before anything reaches a regulator-facing decision — isn't a one-bank trick. Between roles, I set out to prove it: not as a slide, as working code. Six agents, one shared controls library, all live on GitHub.

What's live

Each repo targets a real regulatory or operational pain point in financial services. None of them are demos — each ships with an evals harness, explicit guardrails, and a README that states the design trade-offs, not just the features.

$ ls github.com/basavarajshepur-lab --status=live
01 model-risk-copilot SR 11-7 / SS1/23 model validation — documentation analysis, compliance-gap scoring, auto-drafted validation reports. Python · Claude · RAG 02 kyc-aml-intelligence-agent Entity screening and adverse-media triage, with SAR narrative drafting gated behind a confidence threshold. Python · Claude 03 aml-copilot AML case triage copilot — surfaces the evidence an investigator needs, never auto-closes a case. Python · Claude 04 kpi-integrity-agent Anomaly detection that checks the data quality behind a KPI spike before trusting the signal — not just whether the number moved. Python · Claude · Streamlit 05 Data-Intelligence-Platform-multiagent Metadata, lineage (sqlglot, 20+ SQL dialects), and DAMA-DMBOK data quality scoring — the pattern every agent above is built from. Python 06 responsible-ai-finserv ConfidenceGate · AuditChain · HITLQueue · DriftMonitor · ExplainabilityWrapper — the shared controls library every repo above imports. Python · Shared library

The architecture

The interesting design decision wasn't any single agent — it was refusing to let five different repos reinvent the same safety controls five different ways. Every agent that makes a decision a human would otherwise have to catch imports the same shared layer.

Fig 01 · Shared controls layer, reused across every agent responsible-ai-finserv · Import graph
AGENTS · CONSUME THE SHARED LAYER model-risk-copilot kyc-aml-agent aml-copilot kpi-integrity-agent Data-Intelligence- Platform-multiagent SHARED LAYER · IMPORTED, NOT REIMPLEMENTED responsible-ai-finserv ConfidenceGate AuditChain HITLQueue DriftMonitor ExplainabilityWrapper blocks low-confidence output · logs every decision with its prompt hash · queues anything high-risk for human review

One controls library, five agents importing it. Fix a gap once — every agent above inherits the fix.

Why this, not another slide deck

Measured impact
6 Production-grade AI agents shipped and live on GitHub.
1 Shared controls library imported by every agent above.
20+ SQL dialects supported by the lineage-parsing agent, with no database connection required.
0 Agents that auto-approve a high-risk decision without a human gate.

A hiring manager can read a bullet point about "responsible AI" on any resume. This is the difference between the claim and the artifact: the code is public, the guardrails are documented in each repo's README, and the pattern is the same one that got sign-off from a tier-one bank's Chief Compliance Officer.

The full picture

Five programmes, one consistent instinct.

Start at the beginning for the trajectory, or jump straight to the programme closest to the problem you're hiring for.

Read case study 01 → All case studies