Making an AI fraud engine a reviewer can trust.
Sentinel is a fraud-detection & risk-review dashboard for credit-union operations teams. It turns an opaque risk score into an explainable, role-aware workflow — so junior reviewers, senior analysts, and compliance leads each get exactly the interface their job needs.
Explore Sentinel, live
The frame below is the actual deployed React app. It's fully interactive — click rows, open the detail panel, filter the queue, and use the “View as” switcher to experience all three roles.
Fraud tools flag transactions. They don't help humans decide.
A credit union processes thousands of transactions a day. A model can flag the suspicious ones — a $48K wire to a brand-new recipient, a login from two cities in two hours — but the score alone creates more problems than it solves for the team that has to act on it.
- !Black-box scores kill trust. “Risk: 75” tells a reviewer nothing about why, so they can't defend the decision — or learn from it.
- !One screen for every job. A junior reviewer, a senior analyst, and a compliance auditor have completely different authority and needs, yet most tools show them the same thing.
- !No safe handoff. Escalations get lost in email; seniors act without the junior's context; nothing is auditable.
- !Edge cases break the flow. When the scoring service is down or two people act on the same case, the UI has no answer.
Design goal
Turn a raw AI score into a trustworthy, role-appropriate decision surface — explainable enough for a junior, powerful enough for a senior, and airtight enough for compliance.
Three roles, three very different jobs
The same transaction means something different depending on who's looking. Permissions aren't a backend detail — they reshape the entire interface for each persona.
Junior Reviewer
Senior Analyst
Compliance Lead
Grounding the design in real fraud-ops behaviour
Before a single screen, I mapped how fraud reviews actually work and what makes AI-assisted decisions succeed or fail — then let that shape the product's non-negotiables.
What I studied
- →The real product language. I analysed Candescent's live product UI and deliberately adopted its light, airy visual system — rather than the dark palette in the written brief.
- →Fraud typologies. I catalogued the fraud patterns credit unions genuinely see — account takeover, structuring, check kiting, impossible-travel card fraud — and turned each into a scoring rule.
- →Explainable-AI heuristics. Research on human-AI trust is clear: people accept model output when they see the signals behind it, not just the number.
Principles that fell out of it
- ✓Always show the “why”. Every score is backed by named signals and a plain-language headline.
- ✓Least privilege by design. If you can't do it, you shouldn't see an enabled control for it — but you should understand why.
- ✓Design the unhappy paths. Service-down, conflicting actions, and empty states are first-class screens, not afterthoughts.
- ✓Match the customer's brand. Feels like it belongs inside Candescent's suite on day one.
From fraud logic to a shipped, interactive product
I treated the build itself as part of the design argument — a legible, explainable system rather than a static mockup.
Model the brain first
Built a deterministic, 11-rule scoring engine so the whole UI reads from stable, explainable data — score, band, signals, anomaly, and confidence.
Define the permission matrix
Mapped every capability to every role. This became the single source of truth that gates both routes and individual buttons.
Design the system, not screens
Authored a design-token layer (color, type, spacing, risk semantics) up front so every component stays consistent and on-brand.
Build the core loop
Queue → detail panel → action → activity/audit. Then layered role-specific surfaces: escalation inbox, insights, audit trail.
Harden the edges
Added stale-data fallback, conflict detection, empty and loading states, and accessibility passes (contrast, focus, reduced motion).
Ship it live
Deployed a real, interactive React app so the case study is the product — recruiters can use it, not just look at it.
| Fraud-detection rule | Signal |
|---|---|
| Large withdrawal · top 5% of history | +30 |
| Impossible travel · 2 cities, under 2h | +28 |
| Loan ID mismatch | +27 |
| Check kiting | +26 |
| New account · under 30 days | +25 |
| Card geo-spread | +24 |
| Unusual time · Structuring · Velocity … | +20–23 |
11 rules sum to a capped 1–100 score. The transaction's actual fraud type — not just the highest-point rule — becomes the human-readable headline.
The choices that make Sentinel trustworthy
Show the reasoning, not just the number
Instead of a bare “75”, the detail panel leads with the fraud type as a plain-language headline, then breaks down each triggered signal and its points — plus an anomaly comparison against the account's own history and a model-confidence read. A reviewer can defend the call and, over time, learn how the model thinks.
Permissions you can see
Juniors see disabled actions with a tooltip explaining why, not hidden ones — so the boundary is legible, not mysterious. Routes, toolbars, and buttons all consult one capability matrix.
Escalation with context
A junior's note travels with the transaction into the senior's inbox, so the decision-maker inherits the reasoning — and every action lands in the audit trail automatically.
A dedicated “AI voice”
A gold gradient is reserved only for model-generated content — insights, smart-escalation, confidence. Users instantly know when it's the AI talking versus plain data.
Designed edge cases
Scoring-service-down dims metrics and drops to a rules-only fallback; concurrent actions surface an “already actioned by…” notice; every list has a purposeful empty state.
Bulk actions
Real reviewers batch. “Approve all under $500” and multi-select actions match how the work is actually done, not one-row-at-a-time toil.
“View as” role switcher
A single control swaps the entire experience between the three personas — so anyone can see all three workflows in seconds, no logins required.
One workflow, tailored to every role
Six connected surfaces make up the product. Each answers a specific question for a specific person.
Dashboard
Five live metrics, a right-rail activity feed, and five purpose-built charts — risk distribution, 7-day trend, top reasons, model calibration, and reviewer performance.
Review queue
Sortable, filterable table with risk chips, search, bulk actions, and a slide-in detail panel — the core loop where juniors and seniors spend their day.
Risk detail panel
Account history, the explainable signal breakdown, anomaly detection, metadata, a notes thread, and role-gated actions — all in one focused overlay.
Escalation inbox · Senior
Everything juniors sent up, with their notes and AI confidence attached, ordered so the most urgent decisions surface first.
Audit trail · Compliance
A read-only, filterable log of every action — who, what, when, why — with one-click CSV export for regulators.
AI insights · Senior / Compliance
Precision, recall, false-positive rate, trending patterns, and confidence calibration — so the humans can judge the model.
What the finished product proves
If I took it further
Usability-test the escalation note flow with real reviewers, add keyboard-first queue navigation for power users, and let a clicked donut segment filter the queue directly.
The core lesson
AI features live or die on trust. The design work isn't the model — it's exposing the model's reasoning, respecting each user's authority, and never leaving them stranded on an edge case.
See it for yourself
The best way to understand Sentinel is to use it. Switch roles, open a high-risk case, read the signal breakdown, and escalate one up the chain.
Open the live app ↗