01 / The narrow harbour
The sea doesn’t stop.
The queue has to.
UPI Fraud Engine
A story about the signals we can review, and the ones we miss. Scroll to enter the harbour.
Follow the current ↓02 / Time moves one way
Only the past
can guide this light.
Transactions arrive with history, not hindsight.
Temporal and behavioural features use information available before the transaction. The frozen model returns a risk score. These moving lights are an illustration, not a transaction log.
03 / The review gate
The harbour
has a capacity.
A 0.5% alert budget is the project’s operating assumption.
The highest-risk fraction enters review. The opening is a visual metaphor, not a score threshold scale or evidence of a real team’s staffing. A risk score is not a fraud verdict.
Budget limits the queue.
It does not make the sea safe.
04 / One evaluation
A bright signal.
A narrow reach.
Held-out evaluation · 85,429 transactions
Precision describes the flagged set. Recall describes the fraud that was caught. At this budget, most fraud is missed. Read both numbers together.
Production XGBoost evaluation · 22 January 2026
05 / Seven different tides
Now let
the days pass.
A separate replay · 1–7 June 2025
Choose a recorded day. The dial moves through real daily precision, recall and alert counts. It does not simulate an untested threshold.
16 alerts / 3,101 transactions
06 / Two records, two questions
The second number
gets its own story.
Replay · 22,071 transactions across seven days
The replay recorded 75.22% precision and 12.13% recall. It is not the held-out test. The replay reports zero budget violations under its policy; daily alert counts use whole transactions.
07 / Outside the light
Keep the darkness
in the picture.
85 caught. 616 missed. Of 701 fraudulent transactions.
Each point here represents one fraudulent transaction counted in the replay. Positions are editorial. The 28 false positives are separate: 113 alerts included 85 true positives and 28 false positives.
85 caught inside. 616 missed outside. Both remain part of the record.
08 / A signal, not a verdict
Help someone
look closer.
Prioritise a review. Leave the verdict to it.
UPI Fraud Engine’s useful contract is a bounded queue with its trade-offs visible. Longer replay and real review feedback remain future work, not claimed operational impact.
Read the engineering story ↗Return to the current ↑Inspect the seven-day record
| Date | Rows | Alerts | Precision | Recall |
|---|---|---|---|---|
| 2025-06-01 | 3,101 | 16 | 68.75% | 10.68% |
| 2025-06-02 | 3,896 | 20 | 75.00% | 12.40% |
| 2025-06-03 | 3,105 | 16 | 100.00% | 12.50% |
| 2025-06-04 | 3,783 | 19 | 57.89% | 11.34% |
| 2025-06-05 | 3,352 | 17 | 100.00% | 13.71% |
| 2025-06-06 | 2,575 | 13 | 84.62% | 19.30% |
| 2025-06-07 | 2,259 | 12 | 33.33% | 5.63% |
Held-out and replay records remain separate. No live transaction scoring occurs in this world.