Kubrainn symbol Kubrainn

IDShield

Fraud & identity sentinel with layered signals

Reasoning Vision Dual

Overview

Multi-layered fraud detection with real-time network analysis

7-Layer Fraud Signal Stack

  • ✓ Device Fingerprinting: Hardware + OS + browser entropy
  • ✓ Liveness Detection: Blink, head turn, smile challenges
  • ✓ Velocity Checks: Applications per device/IP/email
  • ✓ Network Analysis: Fraud ring detection with graph ML
  • ✓ Document Verification: OCR + tamper detection + Vision AI
  • ✓ Behavioral Biometrics: Typing rhythm, mouse patterns
  • ✓ Watchlist Screening: PEP, sanctions, adverse media

Key Features

  • ✓ 3D Network Visualization: WebGL-powered fraud graph with Three.js
  • ✓ Vision-First Approach: Document + Selfie analysis in one pass
  • ✓ Escalation Automation: Auto-route high-risk cases to L2 review
  • ✓ Real-Time Scoring: < 3s end-to-end latency
  • ✓ Explainable Output: Every decision backed by signals
  • ✓ PII Masking: Automatic redaction in logs & exports
450-650%
Annual ROI
93%
Fraud Block Rate
0.2%
False Positive Rate
< 3s
P95 Latency

Tech Stack

Vision + Reasoning engines with graph ML

Technology Architecture

Vision Engine
  • Document extraction (Aadhaar, PAN, DL)
  • Liveness detection (3 challenges)
  • Face matching (selfie vs ID)
  • Tamper detection (shadows, fonts, pixels)
Reasoning Engine
  • Signal aggregation (7 layers)
  • Risk scoring (0-1000)
  • Fraud ring detection
  • Explainable output with chain-of-thought
Graph ML
  • Neo4j for fraud networks
  • Community detection (Louvain)
  • PageRank for entity importance
  • 3D visualization (Three.js + WebGL)

Data Flow

End-to-end fraud signal pipeline

Fraud Signal Pipeline
flowchart LR A[Client Upload] --> B{Doc + Selfie} B --> C[Vision Engine] B --> D[Device FP] C --> E[OCR + Face Match] D --> F[Velocity Check] E --> G[Reasoning Engine] F --> G G --> H{Risk Score} H -->|Low| I[Auto-Approve] H -->|Medium| J[Manual Review] H -->|High| K[Auto-Reject] G --> L[Graph DB] L --> M[Fraud Network] M -.Update.-> G style C fill:#8B5CF6 style G fill:#A3E635 style M fill:#E10600

Signal Catalog

7-layer fraud detection framework

Signal Layer Sequence
sequenceDiagram participant Client participant FP as Device FP participant Vision as Vision AI participant Velocity participant Graph as Network Graph participant Reasoning Client->>FP: Browser fingerprint FP->>Reasoning: Device signals Client->>Vision: Doc + Selfie Vision->>Reasoning: OCR + Liveness Velocity->>Reasoning: Velocity metrics Graph->>Reasoning: Network risk Reasoning-->>Client: Risk score + reasons

Benefits

For NBFCs and applicants

NBFC Gains

  • 93% fraud block rate (vs 78% manual)
  • 0.2% false positive rate (vs 3-5% industry)
  • 450-650% annual ROI
  • Auto-escalation reduces L1 workload by 60%
  • 3D fraud network visualization for investigators
  • Explainable decisions for audit trails

Borrower Gains

  • Faster approvals (< 3s vs 20min manual)
  • Lower rejection rate (fewer false positives)
  • Seamless liveness checks (3 simple challenges)
  • PII protection (automatic masking)
  • Transparent reasons if rejected
  • Single upload (doc + selfie)

Interactive Demos

Explore fraud detection in real-time

Escalation Simulator

Adjust fraud signals to observe risk level, escalation tier, and decision

Signal Contributions
Fraud Risk: LOW
Tier: Auto-Approve
Decision: PASS
View JSON Output 
{
  "fraud_risk": "LOW",
  "escalation_tier": "AUTO_APPROVE",
  "decision": "PASS",
  "reasons": ["Strong device trust", "High liveness score"],
  "signals": {
    "device_trust": 85,
    "liveness": 92,
    "velocity_flags": 0,
    "network_risk": 10,
    "doc_quality": "high"
  }
}

Mode: MOCK

Toggle between offline simulations (MOCK) and live API calls (LIVE)

FAQ

How does IDShield differ from traditional KYC?
Traditional KYC is binary (pass/fail) and manual. IDShield provides a continuous risk score (0-1000), detects fraud rings through network analysis, and auto-escalates high-risk cases. Vision AI enables document tamper detection that human reviewers often miss.
What is the 3D fraud network visualization?
IDShield builds a graph database (Neo4j) connecting applicants by shared signals (device, IP, email, phone). Investigators can explore this network in 3D (Three.js + WebGL) to identify fraud rings. Nodes are color-coded by risk level, and edges show connection strength.
How are false positives minimized?
IDShield uses a 7-layer signal stack with conservative thresholds. No single signal causes auto-rejection. The Reasoning engine weighs all signals and requires multiple red flags to escalate. This achieves 0.2% false positive rate vs 3-5% industry average.
What documents are supported?
Vision engine supports Aadhaar, PAN, Driving License, Voter ID, and Passport. OCR extracts text, tamper detection checks for digital manipulation, and face matching compares selfie to ID photo. All PII is automatically masked in logs and exports.
How does device fingerprinting work?
IDShield collects ~50 signals: browser version, screen resolution, timezone, installed fonts, canvas fingerprint, WebGL renderer, battery level, etc. These create a unique device ID that persists across sessions. Velocity checks track applications per device to detect farming attacks.
Is IDShield RBI compliant?
Yes. IDShield follows RBI Fair Practices Code 2003, DPDP Act 2023, and Aadhaar Authentication Guidelines. All decisions include explainable reasons, PII is masked, and audit trails are exportable. No biometric data is stored permanently.
← RiskCore ↑ Technology Overview Collective →
Try Live Demo