Fraud is a network.
Score it like one.
A per-transaction model sees one order at a time. Trust Graph links buyers, sellers, delivery partners, devices, IPs and addresses into one actor graph — then explains what it found in language a human reviewer and an accused seller can both act on.
Refund abuse, seller–buyer collusion, fake delivery scans and rating inflation are coordinated across accounts. Scored one transaction at a time, each order looks ordinary. Meanwhile every false positive freezes someone's income, so aggressive blocking is not an option.
The approach- Deterministic triage score over transaction, device and velocity features.
- Classical graph solver over shared identifiers to surface collusion rings.
- LLMs only where language and judgement are needed — never for scoring.
- Graduated, time-bound, appealable remediation behind a precision gate.
Success metrics we hold ourselves to
Fraud loss avoided
Measured against confirmed-fraud labels on a held-out slice
Precision ≥ 95% for hard actions
Enforced in code — blocked actions route to a human
Median time-to-resolution
SLA clock on every case, overdue reviews escalate
Action-rate parity
Computed per seller-size and partner cohort, published on the dashboard
Five cooperating agents, not one mega-prompt
Each step runs on the cheapest engine that can do the job. Every run is logged with its tier and estimated cost, so cost-per-decision on the dashboard is measured, not asserted.
| Agent | Job | Tier |
|---|---|---|
| Triage scorer | Deterministic feature rules over transaction, device and velocity features. | No LLM · ₹0 |
| Graph analyst | Union-find ring detection, reciprocal loops, rating inflation, POD anomalies. | No LLM · ₹0 |
| Evidence explainer | Turns numbers and graph structure into narrative for both audiences. | Cheap LLM |
| Remediation planner | Picks the graduated action, sets time bound + SLA, checks the precision gate. | Cheap LLM |
| Self-check reviewer | Audits high-stakes actions against fairness and livelihood guardrails. | Reasoning LLM |
- Precision gate. Suspensions and payout freezes are unavailable below 95% measured rule precision; the UI says why and routes to a human.
- Livelihood. Every income-affecting action carries an expiry, an appeal link and an SLA deadline. Overdue reviews escalate in the queue.
- Auditability. Append-only evidence log per case — no updates, no deletes — written in reviewer-readable language.
- Data residency. PII lives in a dedicated table with a documented India-region processing boundary and investigator-only access.
The classical scorer and graph solver handle the overwhelming majority of decisions at zero model cost. LLM tiers are invoked per case, on demand, by the investigator — and the reasoning tier only on high-stakes cases.
- Real dataset ingestion (IEEE-CIS / Elliptic CSV upload and scoring).
- Live API keys for AbuseIPDB, GSTIN verification and notifications — the adapter layer is already in place, so no call sites change.
- A trained ML model; today's scorer is a transparent feature-weighted model over seeded labels.