“Blocked. Card ending 4021, merchant in Ho Chi Minh. Explain if I was wrong.”
Watcher watches every card transaction, UPI push, account login, device change, and beneficiary addition in real time. Median verdict latency 164ms across 2.3M decisions this quarter; every one of them signed by the Courtroom, hashed into an append-only audit chain, reviewable by the user and regulator alike. She has blocked 1,203 fraud attempts in the last 24 hours and logged every decision, including the false positives, for later review.
Three illustrative cases. Examples of the work, not live data. Each is the kind of decision Watcher closes on a typical week.
A Chennai schoolteacher received an 'urgent' WhatsApp message from her 'bank manager' at 9:42pm. She began typing the OTP into a lookalike page. Watcher's lookalike-URL model caught the domain change in 340ms and blocked the transaction attempt before submit.
A card-not-present charge from a legitimate merchant whose checkout had been compromised. Watcher matched the merchant's transaction fingerprint against a tampered pattern the Observatory had just flagged.
A synthetic-identity loan application with paperwork that looked clean. Watcher cross-referenced phone tenure, device graph, and salary pattern with Saver and Filer.
24/7 — but the busiest hours are the loneliest. Most blocks happen between 23:00 and 05:00.
See the full 24-hour cycle →Transactions, statements, news, SMS — read into the Newsroom and the Archive.
Watcher drafts a proposal. The Courtroom argues both sides. Sub-200ms.
Signed verdict + SHAP trace + counterfactual ladder out to Saver or back to you.
Hash-chained, replayable, never deleted. DPDP §12 erasure cascades clean.
Watcher closes some decisions on her own. Others require a hand-off to one of the other six. The mesh is public — here are the routes Watcher uses most.
0.08% of the time, measured against user appeals. Roughly one in every twelve hundred blocks is a false positive. I log every one, we retrain on the fresh labels weekly, and the number has been coming down. It will never be zero.
I step down a rung. If a transaction is on the edge, I do not block — I ask. A push notification with the merchant, the amount, and a one-tap approve. If you do not answer in 90 seconds and the risk is above a hard floor, I hold. If you answer, I learn.
They try. The six models are rotated and retrained on different schedules for exactly this reason. A pattern that defeats the lookalike-URL model still has to defeat the device graph and the merchant fingerprint. The council vote is the defence.
Harder. My job is not to judge intent; it is to flag anomaly. If a user suddenly sends ₹2 lakh to a new beneficiary at 2am and the device-graph says someone else is in the app, I hold. If the user is awake and typing normally, I let it through and log the size. Persistent patterns I escalate to a human reviewer.
Same specialist. Same Courtroom-signed decisions. Available via our B2B API — priced by verdict, audited by default.