Without change, analysts keep triaging noise while churn risk stays invisible until the customer is already gone.
Chapter 2 · Role and boundaries
Two users, one screen.
I owned
User modelling: personas and empathy map
SWOT and product positioning
Wireframes for every screen
UI guide: type scale and colour system
Final dashboard, data sources, churn, users and insights screens
Boundaries
Proof-of-concept with limited field validation
Desktop only in this version (no mobile responsiveness)
Depends on clean, accurate data streams
Two users, one screen
Fraud and retention look like different jobs, but both need the same thing: see risk early and adjust how sensitive the system is without filing a ticket.
Tap a card to flip it ↻
Chapter 3 · Key decisions and trade-offs
Every change shows its cost first.
Decision 1 of 4
Show the cost of a threshold change before it happens
Empathy map: “I need control over how alerts behave.”
Options on the table
Why this one
Analysts tune sensitivity every day. The real question is what a change costs, so the preview shows fewer alerts next to the fraud that would slip through.
What I gave up
A backtest replays last month, so it can't see new fraud patterns. The screen says so, and every change keeps a reason and a revert path.
Decision 2 of 4
Separate who can change detection from who approves it
Hard question: a slider lets anyone loosen fraud detection.
Options on the table
Why this one
Loosening detection is the riskiest action in the product. Splitting the roles means no single person can loosen rules and hide it.
What I gave up
Larger changes wait for a lead. Time-limited approvals ("apply for 7 days") keep urgent changes possible.
Decision 3 of 4
Fraud and churn in one dashboard
SWOT strength: unified view; both personas need early risk signals.
Options on the table
Why this one
Fraud and churn signals overlap in the same subscriber data. One screen lets both teams see the same customer risk picture.
What I gave up
More on the first screen. The landing view shows only top KPIs, and depth lives in dedicated Churn and Insights pages.
Decision 4 of 4
Designed for non-technical operators
Pain point: “Hard to interpret raw data.”
Options on the table
Why this one
Retention leads and managers need to act, not query. Visuals make anomalies obvious at a glance.
What I gave up
Advanced users may want raw access; noted as a known gap in the SWOT.
Try it
Try it: loosen the SIM-box threshold.
This is the decision the whole product is built around. Drag the slider the way an analyst would and watch what the preview tells you before anything goes live.
LooserStricter
Live setting: 6. You're previewing 6.
Alerts per day40no change
Fraud cases missed per day3no change
False positives62%of alerts
No change to submit.
Illustrative numbers, shaped like the prototype's sample data. In the product the preview comes from a backtest on last month's alerts.
Wireframes: users overview, fraud detection, trends and network monitoring.
What my own critique of version 1 found, and what changed
01
The home screen spent its top third on a welcome card and a banner
It now opens on the alert queue, ranked by risk and waiting time
02
Numbers without units ("Fraud detection rate: 123")
Every figure has a unit and a comparison: false-positive rate, time to action, fraud blocked in rupees
03
Sliders changed detection with no feedback
A backtest preview shows alerts per day, missed fraud and false positives before submitting
04
Anyone could loosen detection
Roles: analysts, Fraud Leads, admins; larger changes need approval and a written reason
Each source shows health, last record and volume; failures explain what they delay
Analyst home: open alerts ranked by risk and age, with bulk actions, the live threshold and who last changed it.Adjusting a threshold: the distribution, the backtest impact, the approval rule and a required reason, with the change history alongside.The Fraud Lead's view: what changes, the expected impact and its limits, and the option to approve for a fixed period.
Users and roles, with what each role can do. Admins manage access but can't change detection.Data sources with health, last record and volume. A failure says which alerts it delays.
Churn risk for the retention lead: the forecast, the reasons behind it and segments ranked by revenue at risk, each with a next action.UI guide: Helvetica Neue type scale, an electric-blue primary and an amber secondary.Chapter 4 · Outcome and evidence
I reviewed every screen against Nielsen's 10 usability heuristics, then walked through the four test tasks below step by step as a new analyst. Findings are rated by how likely they are to cause a wrong decision.
7 findings · 6 fixed in the screens · 1 open
HighThreshold preview · match with the real worldFixed
"Missed fraud 0.6% → 2.1%" reads as a small change, but it's three times as many cases.
The preview now leads with cases per day (3 → 9) and shows the multiplier.
HighApproval · error preventionFixed
The 7-day time limit sat inside the Reject card, so a lead could set it and then press the wrong button.
Approve and Reject are separate cards; the time limit belongs to Approve.
HighWalkthrough · why are top-up alerts late?Fixed
A failed data source was only visible to admins. Analysts saw fewer alerts and no reason.
Home shows a delay banner that links to the data source.
MediumHome · consistencyFixed
The card said "Detection threshold", but thresholds are set per fraud pattern.
The card names the pattern (SIM-box).
MediumUsers · securityFixed
An analyst with two-factor sign-in off could still block subscribers.
Blocking is disabled until two-factor is on, and the table says so.
MediumAlert queue · efficiencyFixed
No way to assign or close several similar alerts at once.
Rows can be selected; a bar offers Assign to me and Close as false positive.
LowHelpOpen
Analysts can't see what drives a risk score.
A "Why this score" panel listing the top signals.
How I'll test it next
Sessions with 3 to 5 fraud or revenue-assurance analysts, on a clickable prototype with realistic data.
Task 1Find the alert you should handle firstPasses if: Picks a high-risk, long-waiting alert without hints
Task 2Cut alert volume without missing much fraudPasses if: Reads the preview correctly before submitting
Task 3Approve or reject a colleague's changePasses if: Can explain what the change will do
Task 4Find out why top-up alerts are latePasses if: Reaches the failed data source within a minute
Results and the changes they lead to will be added here.