Credit Risk 360
How a $12.8B commercial lending book cut modeled default exposure 31% in one year.
An enterprise credit-risk platform that gives lending teams real-time portfolio sensing — with a GenAI layer that turns live risk data into plain-language narratives and explainable, auditable recommendations.
Explore the live dashboard
A working slice of the product. Switch the time range to watch the trend, sector cohorts, and the GenAI risk narrative respond in real time.
Credit Risk 360 · Live Demo
Total Portfolio · as of May 2025
Default Probability Trend
1Y · weighted PDPD by Sector Cohort
1Y PD %AI Risk Narrative
GenAIPortfolio risk has drifted into a high band, driven by rising default probabilities in Consumer and Construction, softer macro conditions, and credit migration in recent vintages. If the 1Y trend persists, modeled expected loss could rise ~$22M (+17%) next quarter. Recommended action: tighten limits on Construction cohorts above the 90th PD percentile and stage proactive provisioning.
The Challenge
Credit officers were accountable for a $12.8B commercial book but were reading it through scattered spreadsheets and overnight batch reports. Defaults surfaced only after accounts had already migrated down the rating scale, so every intervention was a post-mortem. The status quo was unsustainable: provisions kept climbing and regulators were asking for evidence of early-warning capability the bank could not produce.
What We Learned
Structured workshops with credit officers surfaced the insight the product turned on: experts reason in cohorts and trajectories, not static scores. A single account's rating matters far less than the direction its vintage is heading. Anything that showed a snapshot was, to them, the wrong question answered precisely.
The Strategy
Build one pane of glass around trajectory, not status. Alerts open into a cohort view, a cohort opens into an account timeline, and every screen carries the direction of travel. A retrieval-grounded LLM reads live portfolio state — PD trends, cohort migration, macro indices — and writes a plain-language risk narrative with a ranked, explainable recommendation. Every generated statement links back to the underlying figures, and officers can accept, dismiss, or annotate each one, feeding a decision log the regulator can read.
How It Was Built
The engagement pushed the design system forward: semantic risk tokens (success / warning / danger), a reusable analytics-card primitive, and a consistent charting language of gradient area trends, sector heatmaps, and cohort bars. Those tokens now drive every data surface in the wider platform. The core flow — alert, cohort, account timeline, AI-recommended action, logged decision — preserves context at every hand-off so an officer never re-derives what the system already knows.
The Impact
Modeled default exposure fell 31% through earlier intervention and review cycles ran 2.1× faster. Analyst adoption rose 38% as the AI narrative removed the blank-dashboard problem. Provisioning shifted from reactive to proactive, and the decision log materially raised regulator confidence in the bank's early-warning story.
What I Carry Forward
Match the interface to the expert's mental model — trajectories, not snapshots — and adoption follows without a training programme. GenAI earns trust only when every claim is grounded in a visible number and every recommendation is reversible.
Behind the work
My scope, the guardrails I worked inside, what the research changed, and how the work travelled from sketch to production.
My role & scope
Design lead, end-to-end: discovery workshops with credit officers, information architecture, the alert-to-decision flow, the risk-narrative interface, and the semantic risk tokens shipped into the platform design system. Worked with 4 engineers, a risk PM, and a model-risk reviewer over 9 months.
Constraints & guardrails
- Model-risk governance: every AI statement had to be traceable to a visible figure before it could ship.
- Legacy batch pipeline — portfolio state refreshed overnight, so the UI had to communicate data age honestly.
- Bank design standards and WCAG AA contrast on dense, colour-coded risk data.
- No net-new training for 300+ officers; the interface had to be learnable in one sitting.
What research changed
- Shadowing 14 credit officers showed they reason in cohorts and trajectories, never single-account scores.
- Analytics showed 62% of alert clicks ended with no logged action — the alert carried no context to act on.
- Usability testing on the first build: 5 of 8 officers could not explain where a score came from.
The pivot
The first direction was a scored watchlist ranked by PD. It tested badly — officers distrusted a number they could not decompose. We rebuilt around trajectory: alert opens a cohort, cohort opens an account timeline, and every AI claim links back to the figure behind it.
UX metric → business impact
Whiteboard alert-to-decision maps and paper cohort flows tested with officers before any UI existed.
Clickable prototype of the cohort drill-down plus annotated states for empty, stale and low-confidence AI output.
Live dashboard with semantic risk tokens, gradient trend charts, sector heatmaps, and the grounded narrative panel.
Retrospective
Instrument the decision log from day one instead of month four — we lost early baseline data that would have made the exposure claim tighter.
In regulated enterprise work, explainability is the adoption strategy. Officers accept AI guidance only when they can decompose it, and stakeholder alignment gets easier the moment every claim points at a number.
“We stopped finding out about deterioration at month-end. The narrative tells us which cohorts are moving and why, and we can defend every call to the regulator.”