threshold / overview
Overview
7 rules live
review ≥ 40 · block ≥ 65
60-second tour. this is a working fraud rule engine running on one day of transactions.
- drag the amber and red lines on the chart: that's your review and block policy.
- switch the dataset to bot attack, then festive spike. watch the same policy succeed and fail.
- open rules to put a rule in shadow, or transactions to see why any payment scored what it did.
Fraud stopped
0
Fraud missed
0
Good users blocked
0
Review queue
0
Score distribution & decision policy
legitimate above the line, fraud below · drag the thresholds
legitimatefraud
Cost of this configuration
Fraud loss (missed)₹0
Good users blocked₹0
Analyst review time₹0
Total₹0
Analyst capacity
Policy advisor simulated
tell it what you're optimising for. it tests every review / block pair against today's transactions and recommends one. you decide whether to apply it.
how this works in production
the search is real: it's plain optimisation, no ai needed for the maths. in production a language model would only translate the goal ("keep friction low during the sale") into constraints, call a
simulate_policy tool over the last 30–90 days of labelled traffic rather than one day, and explain the trade-off. the change would go through shadow mode and a named approver, never straight to live. here the goal parsing is keyword-based.Rule stack
points add up to a 0–100 score. shadow rules are evaluated and logged but don't change the score.
| Rule | Condition | Status | Points | Hits | Precision | To queue | Owner |
|---|
precision = share of a rule's hits that were really fraud. a rule under 40% is flagged amber: it's mostly annoying good users. click any rule for its shadow backtest.
show ground truth
| Transaction | Time | Amount | Account | Signals | Score | Decision | Label |
|---|
Queue load
Your decisions
reviewed 0correct 0wrong 0avg time —
you're the analyst. open a case, read the evidence, ask the assistant if you need to, then decide. you'll see if you were right.
Manual review
sorted by score, highest first
Cost assumptions
these drive the cost panel. they're mine, and they're arguable. change them.
₹
lost margin + a support contact + some chance they churn.
₹
≈ a ₹7L/yr analyst, fully loaded, over ~220 working days.
min
manual reviews typically run 3–10 minutes.
people
each works an 8-hour shift on the queue.
%
share of fraud in the queue a reviewer correctly stops.
Audit log
every change to rules, policy and assumptions, with who and when.