One governance layer across 80+ models: credit risk, fraud, and trading.
The Challenge
Capital One's model inventory had grown faster than its governance infrastructure. Credit underwriting, fraud detection, and algorithmic trading models each had their own documentation standards, separate validation queues, and siloed monitoring dashboards. When regulators requested evidence of ongoing performance monitoring for ten specific models, the team spent three weeks manually assembling records from spreadsheets, email threads, and disconnected pipeline logs. The OCC examination was six months away, and a second examination cycle under the same conditions was not survivable.
Fraud models presented an additional problem. Real-time scoring at transaction speed meant that any governance intervention had to work without adding latency but the existing process required validators to review batch-exported logs days after inference, making behavioral drift invisible until it affected loss rates.
The Solution
The platform was connected to Capital One's existing model registry, Feast feature store, and CI/CD pipelines without migrating any infrastructure. From day one, every model promotion automatically generated a structured model card, captured data lineage from feature engineering to scoring output, and assigned a risk tier based on the firm's own classification rules encoded in the policy engine.
Fraud models were instrumented for real-time drift detection: data drift on incoming transaction features, prediction drift on score distributions, and fairness drift on approval rate disparities by demographic segment. Threshold breaches trigger an automated alert before they appear in loss reporting. Validators now spend their time on the exceptions the system escalates, not on building the dossiers themselves.
Continuous Monitoring at Transaction Speed
The fraud scoring stack processes hundreds of thousands of transactions per hour. Governance instrumentation was embedded at the inference layer, not bolted onto a downstream batch process, so drift signals are captured in real time without adding measurable latency. When feature distribution shifts exceed the configured threshold, the relevant model is flagged for expedited human review before the next business day.
SR 11-7 Evidence as a Pipeline Output
The OCC's model risk guidance requires ongoing performance monitoring, clear documentation of model purpose and limitations, and an independent validation function. Each of those requirements maps to a specific output the platform generates automatically: performance benchmarks from CI runs, model cards authored from registry metadata, and a separation of the build and validation roles enforced at the approval gate. Exam preparation became a matter of running a report, not reconstructing a history.
Fairness Drift Across the Credit Portfolio
Capital One's consumer lending models carry fair lending obligations under ECOA and the Fair Housing Act. The bias detection module tracks approval-rate disparities across protected classes at inference time, not just at training time. When a model's fairness metrics drift outside pre-approved bounds, it escalates automatically to the model risk committee, creating an auditable record of detection, escalation, and remediation that satisfies both internal policy and regulatory expectation.
Results
Faster drift detection allowed model updates to ship before score degradation reached business impact.
Complete SR 11-7 evidence package assembled from existing pipeline data, no manual reconstruction.
Every production model carries a documented classification, lineage trace, and active drift monitor.
Model performance records that took days to assemble are now exported on demand.
See what Thndr AI can do for your team
Talk to our team about your specific AI governance challenges.
