Vol. II · NPCI Hackathon 2026Blue Team · UPI Fraud Detection

Privacy-Preserving Collaborative Intelligence

Varaksha.

UPI Fraud Defense Network — Version II

Rust · Random Forest · IsolationForest · NetworkX · Bhashini
Intelligence Briefing

Threat Magnitude

₹1,750Cr

Lost to UPI Fraud

FY 2024

— RBI Annual Report 2024

Transaction Scale

640M

Daily UPI Transactions

Peak Active Volume

— NPCI

Model Performance

94.4%

Random Forest Accuracy

75K rows · 4 real datasets

— Varaksha V2 · trained Mar 2026

Gateway Performance

<10ms

Rust DashMap Cache Latency

P99 Response Time

— Varaksha Gateway · Criterion.rs

Accessibility Imperative

44%

Indians Lack English Proficiency

Language Barrier

— India BRICS Language Survey

Architecture Thesis

Decoupling Intelligence from the Payment Path

The Constraint

Every UPI payment races through a 15-second settlement window. Conventional fraud scoring cannot share that critical path with heavy ensemble models, cross-bank graph traversals, and consortium lookups — the latency cost alone would collapse transaction throughput at NPCI scale.

The Verdict

The Rust gateway issues a binary ALLOW / FLAG / BLOCK verdict in under 10 ms from consortium-built DashMap risk intelligence. The ML ensemble and graph agent run asynchronously, entirely off the critical path — continuously enriching the cache, never once blocking the transaction.

Five-Layer Stack — Varaksha V2async · off-path · consortium

L1

ML Engine

RF-300 · IsolationForest · SMOTE

Active

L2

Rust Gateway

DashMap · SHA-256 · <10ms

Active

L3

Graph Agent

NetworkX · Fan-out · Cycles

Active

L4

Alert Agent

LLM · Bhashini NMT · TTS

Active

L5

Dashboard

Real-time Verdict Stream

Active