Every rule, feature and threshold in this specification is anchored in measurable signal from two operational mule accounts: an upstream UPI-collection account and a downstream NEFT-dispersal account, both fed by the same source. These are not hypothetical thresholds — they are observed.
MuleHunter.AI is not a single model. It is a stack — onboarding intelligence, deterministic rules, behavioural ML, graph reasoning and federated network signals — collapsing to one continuous score per account, updated every transaction.
The model never sees a transaction directly. It sees per-account, per-window aggregations across six families. Each family was instrumented to capture a specific class of mule behaviour observed in the anchor cases.
VOL/ volume, count, peak-velocity, burst-durationTMP/ hour-of-day skew, dormancy-then-burst, inter-arrivalCNT/ sender-count, sender-bank-spread, repeat-ratioDSC/ descriptor entropy, random-token %, phone-VPA %BAL/ retention ratio, max-balance, velocity multipleOUT/ destination concentration, ticket-band, sub-threshold sweepL1 deterministic rules. Each rule traces back to a specific signal in the anchor SBI cases. Severity, logic, window and recommended tier are versioned in the rules registry; thresholds are A/B tested monthly against the I4C-confirmed mule ledger.
Mule behaviour is not a single shape. The pass-through funnel, the dormant-then-burst, the sequential mule fleet — each has a different temporal signature, sparsity and counterparty topology. MuleHunter.AI uses four specialised learners, each optimised for one failure mode, fused at the score layer.
| Model | Specialty | Input | Why this learner | Weight in CMAI |
|---|---|---|---|---|
| M1 · XGB-RANK | Gradient-boosted ranker | 142 tabular features over 24h / 7d / 30d windows | Robust to mixed scales, handles missingness, ships with SHAP explanations — required for explainable banking AI. | 0.30 |
| M2 · T-CNN | Temporal CNN for burst signatures | Per-minute credit/debit count & amount tensors, 7-day window | Convolution detects the dormancy-spike-collapse pattern that tabular features blur. Picks up the 5-hour burst even if smeared across calendar dates. | 0.20 |
| M3 · ISO-NOV | Isolation forest novelty | Account-embedding vector vs historical population | Catches new mule typologies the rules & supervised models have never seen. Important: typologies evolve faster than label data arrives. | 0.15 |
| M4 · GNN-PROP | Graph neural net · GraphSAGE | Sender ↔ account ↔ beneficiary bipartite graph across participating banks | Surfaces co-mule rings, sequential clusters and shared-source funnels. Single-account models cannot see this — by definition, the signal lives between accounts. | 0.25 |
| M5 · RULES | Deterministic L1 firings | R-01 … R-24 boolean vector | Forensic floor. Even if every ML model fails or drifts, a single critical rule (e.g., R-24 I4C hit) can still drive the account to T4. | 0.10 (with veto) |
Every account in every participating bank carries a continuous CMAI score, refreshed on every transaction. Five bands map to four intervention tiers. The score is the only artefact the host bank's core needs to consume — everything else is auditable detail.
A trace of how Account A's CMAI evolves transaction-by-transaction through the burst window. By minute 12 the account would have crossed 85 → T4 lien. By minute 14 the burst is halted. Recovered ₹6.97 Cr of ₹6.98 Cr.
| T+min | State | Rules firing | CMAI | Action |
|---|---|---|---|---|
| T+00 | First 42 credits (avg ₹287, 21 sender banks) | R-04, R-10 | 34 | T1 · watch |
| T+03 | ~520 credits, dormancy break detected | R-04, R-10, R-15, R-13 | 52 | T2 · soft friction on outflow |
| T+06 | 1,800+ credits, peak 4.7 tps, ₹500 dominance | R-01, R-02, R-05, R-11 | 68 | T2 → step-up auth armed |
| T+09 | First RTGS outflow attempt — ₹4,12,500 to a low-history beneficiary | + R-08, R-23, R-07 | 79 | T3 · outflow quarantined, callback queued |
| T+12 | GNN propagation picks up shared upstream with Account B (already at CMAI 91) | + R-18, R-19, R-09 | 96 | T4 · debit freeze, 1601 PD lien |
| T+14 | Bank webhook acks freeze; further credits accepted (legally required) but no outflow possible | frozen | 96 | STR filed · I4C feed updated |
A T4 freeze on a false-positive harms a real customer. A T1 watch on a real mule loses ₹6 Cr. Tier calibration is the most reviewed part of the system — monthly re-tuning against I4C-confirmed and customer-disputed labels.
MuleHunter.AI is an infrastructure-level utility, not a per-bank tool. A confirmed mule at SBI updates the model that scores accounts at HDFC, ICICI, PNB and 50+ other participating institutions, with mule identifiers contributed to the I4C Suspect Registry.