Ethical AI/ML governance & compliance for a leading financial firm

Challenge

The firm had dozens of ML models in production and many more in pre-production, each owned by a different team. Regulators and the board wanted assurance that:

• Every model in production was inventoried, tiered by risk, and traceable to an owner
• High-risk models had documented fairness, robustness, and explainability evaluations
• There was a clear audit trail from data → model → decision → outcome
• Governance did not become a bottleneck that pushed teams to ship outside the process

Approach

VerticalServe’s POD designed a governance platform that met the controls without weighing down delivery:

• Model inventory capturing every production and pre-production model with owner, business purpose, data sources, and downstream consumers
• Risk tiering framework aligned to regulator guidance (e.g., SR 11-7) with clear controls for each tier
• Automated evaluation harness for fairness (disparate impact, equalized odds), robustness, drift, and explainability
• Approval workflows integrated with the firm’s existing GRC and change-management tooling — not a parallel system
• Audit trail linking training data lineage, model versions, evaluation runs, approval records, and production decisions
• Self-service onboarding for data-science teams: register a model, attach evaluation artifacts, and request review — in minutes

Outcome

The firm went into its next exam with a single defensible source of truth for every production model. Data-science teams reported a noticeable drop in “governance friction” because the workflow was integrated into how they already worked. The platform now governs hundreds of models across multiple business units.

Impact

Stack

Python MLflow Great Expectations Insight Catalog Insight Prompts (governance patterns) GRC integration Snowflake Tableau OpenLineage

Read the full long-form on Medium