Engagements that moved the needle.

Context

A top-10 global investment bank had built a portfolio of 40+ AI models across trading, credit risk, and client services — with no unified governance framework, inconsistent data lineage, and growing regulatory scrutiny from the SEC and FCA.

Challenge

Establish an enterprise AI governance program that satisfied regulators in 12 jurisdictions, addressed model explainability requirements under the EU AI Act, and enabled the bank to continue deploying AI at pace without accumulating compliance risk.

Our Approach

We designed a tiered AI risk classification framework aligned to the EU AI Act and SR 11-7 guidance, built a model registry with automated lineage tracking, and embedded privacy-by-design principles into the model development lifecycle. A cross-functional AI Risk Committee was established with board-level reporting.

Context

A national health network was piloting AI-assisted diagnostics and clinical decision support across 200+ facilities. Patient data was flowing into third-party model vendors with no standardized data processing agreements, creating material HIPAA exposure and eroding clinician trust.

Challenge

Build a clinical AI governance framework that protected patient privacy, satisfied HIPAA and state-level biometric data laws, and gave clinical leadership the confidence to scale AI-assisted care without regulatory or reputational risk.

Our Approach

We conducted a full data flow audit across all AI vendor relationships, renegotiated data processing agreements to enforce de-identification and purpose limitation, and designed a federated learning architecture that enabled model improvement without centralizing raw patient data. A clinical AI ethics board was established with patient advocate representation.

Context

A Fortune 100 retailer had built its personalization engine on third-party cookie data. With the deprecation of third-party cookies and tightening CCPA and GDPR enforcement, the entire AI-driven personalization stack was at risk — representing $340M in attributed annual revenue.

Challenge

Rebuild the personalization and AI recommendation infrastructure on a consented first-party data foundation — without degrading model performance or disrupting the customer experience during the transition.

Our Approach

We designed a consent-first customer data platform architecture, built a data clean room capability for privacy-preserving partner data collaboration, and retrained the recommendation models on first-party behavioral signals. A progressive consent program was launched to rebuild the consented data asset at scale.

Context

A national property and casualty insurer was using ML models for underwriting and claims triage. State insurance regulators in 8 jurisdictions had begun requesting model documentation, and internal audits had surfaced potential disparate impact issues in two underwriting models.

Challenge

Remediate fairness issues in live underwriting models, build an explainability framework that satisfied state regulatory requests, and establish ongoing model monitoring to detect drift and bias before it created regulatory or legal exposure.

Our Approach

We conducted a disparate impact analysis across all protected classes for the flagged models, redesigned the feature engineering pipeline to remove proxies for protected characteristics, and implemented SHAP-based explainability reporting for regulatory submissions. An automated model monitoring system was deployed with fairness metric dashboards reviewed monthly by the Chief Actuary.