Financial institutions were managing thousands of AI models across risk, fraud, and compliance — with spreadsheets, disconnected tools, and no way for non-technical stakeholders to audit what the models were actually doing. I led the 0-to-1 design of ModelEdge, an enterprise platform that made opaque AI decisions contestable, traceable, and governed. It remains the most direct evidence I have for a conviction: explainability is a design problem before it is a machine-learning one.
Can a compliance officer audit an AI model’s decision without reading code — and would banks govern models differently if they could?
Lead product designer, 0-to-1. Built and led a five-designer team through 145 interviews and 62 wireframe tests; owned the explainability layer end to end.
1,520+ models under governance; development cycles cut 50–67%; tier-1 pilots at Morgan Stanley, UBS, and Barclays; 491% projected revenue growth.
Since 2008, financial institutions have paid over $321B in regulatory fines. By the time I started this project, banks were spending $100B annually on compliance — and the problem was getting worse, not better. AI models were being deployed across fraud detection, anti-money laundering, credit risk, and transaction monitoring, but the governance infrastructure hadn’t kept up.
The reality on the ground: model development took 6–12 months. Validation took another 6 months. Approval workflows ran on email. AML models had 75–90% false positive rates, and nobody could explain why. Then the LIBOR transition hit — $350 trillion in financial contracts worldwide requiring model updates — and the whole system buckled.
Through 145 interviews and 109 hours of direct customer interaction with stakeholders across large banks, regional institutions, and fintechs, we identified the core problem: there was no single place where a risk manager could see all their models, understand their behavior, and trace why a decision was made.
“Manual tuning, extract data, run in several different environments. Any type of tuning you can do in a singular solution is a huge benefit.”
The design challenge was layered: we had to serve model developers who think in Python and R, risk managers who think in compliance frameworks, and executives who think in portfolio risk. The same platform, three fundamentally different mental models. The working answer — role-based progressive disclosure: one system of record, multiple lenses into it — held up across 62 wireframe-testing interviews.
A single dashboard managing 1,520+ models with real-time status tracking, risk tier visualization, and enterprise-wide metadata. Cloud-based centralized repository with standardized documentation — replacing the fragmented spreadsheets and disconnected tools that risk teams had been cobbling together.
Rule-based approval routing that cut approval time from one month to roughly two weeks. Integrated code workbench (Model Studio) supporting Python and R with comprehensive audit trails. One-click compliance report generation with predefined templates — turning documentation from a bottleneck into an output.
Dashboards with trend analysis, anomaly detection, and optimization insights that let risk managers see when a model was drifting before it became a compliance issue. Calendar-based heat maps for proactive resource planning across validation cycles.
The interface that made the whole platform matter: compliance officers and risk managers could trace data lineage, audit model behavior, and flag decisions made outside acceptable parameters — without needing to read code. This was AI explainability as a design problem, not a machine learning problem.
Data science wanted SHAP values and feature importances exposed everywhere. For a compliance officer, an unexplained explanation is worse than none. We tiered disclosure — plain-language rationale first, statistical depth on request.
Natural-language compliance reports tested brilliantly in demos — and we could not validate their accuracy to a regulatory standard. Cut from v1. A tool that is right 95% of the time is unusable in a domain where the 5% gets you fined.
Engineering costed continuous monitoring at two quarters. Daily batch analysis with anomaly alerts covered every risk case our 145 interviews had surfaced. We shipped the 20% that carried the 80% — and banked the two quarters.
Model development time cut from 6–12 months to 3–4 months. Validation from 6 months to 2–3 months. Approval workflows from 1 month to ~2 weeks. End-to-end cycle time improvements of 50–67% across every stage.
Platform managing 1,520+ models with 546 active assets across tier-1 institutions including Morgan Stanley, UBS, and Barclays. High-probability pilot pipeline with Citi, Capital One, Valley National Bank, and Ally Financial.
491% projected revenue growth from $3.33M to $19.7M. Successfully positioned within the $1.2B serviceable anti-financial-crime software market targeting 750 regional banks and 210 fintech providers.
Built and led a 5-designer team through a rigorous research program: 33 pain point interviews, 50 storyboard validation sessions, 62 wireframe testing interviews. Established design as a core competitive advantage in PwC’s product organization.
“We didn’t build a better dashboard. We built the interface between a model’s decision and a human’s ability to contest it. Years on, I think that remains the design problem that matters most in AI — and it is still mostly unsolved.”
This has been my working question since 2018. Happy to compare notes.
Get in touch