Zarelva · Fraud Intelligence Framework

Decision-Layer Risk Engine

How raw platform events become fraud decisions — and how they continuously improve over time.

Read left to right, top to bottom·Data flows in·Signals are built·Engine decides·Humans tune

01 · Platform & Data Inputs
🪪 Onboarding & KYC Identity, docs, bureau
📱 Logins & Devices Sessions, fingerprints
💸 Transactions & Payments UPI, BNPL, disbursements
⚠️ Disputes & Collections Chargebacks, NPA signals
🤖 AI & Agent Actions Autonomous flows, LLM calls
Raw events from your app — every touchpoint that carries a fraud signal
transforms into signals
02 · Layered Fraud Signal Library
01 Identity & KYC
document mismatch synthetic ID patterns KYC vendor risk codes bureau thin file
02 Access & Device
device fingerprint IP / geo mismatch emulator / root signals session anomalies
03 Behaviour & Velocity
rapid-fire applications typing / mouse patterns time-on-form outliers repeat failure loops
04 Transaction & Monetary
amount outliers merchant / category risk repayment history gaps disbursement timing
05 Network & Ecosystem
shared devices / IPs shared bank accounts link analysis rings mule cluster signals
Signals that actually describe fraud risk — not just raw data
feeds into engine
03 · Decision-Layer Risk Engine
Core Decision Engine
Decision-Layer Risk Engine
📏
Rules EngineHard cutoffs, threshold logic, conditional blocks
📊
ScorecardsWeighted signal combination, composite risk scores
🧠
ML Models (optional)Pattern detection, anomaly scoring at scale
📋
Override Logic & Audit TrailEvery decision documented and defensible
Agent Module
Agent Risk Engine
Prompt injection detection
Delegation chain abuse
Non-human velocity signals
Autonomous flow overrides
Attaches to Behaviour & Network layers. For AI-agent and autonomous decision flows.
Where decisions really happen — approve, review or decline logic lives here
produces outcomes
04 · Decision Outcomes
✅ Low Risk Approve

Signals below threshold. Transaction clears automatically. No friction added.

🔍 Needs Review Review

Signals elevated. Routed to fraud team or step-up verification before proceeding.

🚫 High Risk / Confirmed Fraud Decline · Block · Escalate

Signals confirm fraud. Transaction blocked, account flagged, case escalated for investigation.

Human Feedback Loop
Fraud & Risk Team
Zarelva FPaaS
Reviews outputs, tunes rules, and feeds new patterns back into the engine.
📄 Weekly Fraud Summary Every Monday · 12:00 IST
Incident count, pattern shifts, watchlist updates, top attack vectors.
📊 Monthly Insight Memo Monthly · PDF + call
Loss trends, rule decisions, top patterns. Board-ready language.
🔧 Rule & Scorecard Tuning Ongoing · version-controlled
Recommendations on what to add, retire, tighten. Living backlog.
🏗️ Architecture Review Quarterly · board report
Controls, data flows, gap analysis. Suitable for investors & auditors.
↑ continuous feedback ↑
Identity signals
Device signals
Behavioural signals
Transaction signals
Network signals
Decision engine (Zarelva)
ZARELVA · Decision-layer, not alert-layer · zarelva.com