Skip to content

TrustScan_Master_Architecture_and_Operations_Guide.md

Latest commit

 

History

History
191 lines (151 loc) · 10.9 KB

TrustScan_Master_Architecture_and_Operations_Guide.md

File metadata and controls

191 lines (151 loc) · 10.9 KB
Raw

🛡️ TrustScan AI — Master Architecture & Operations Guide

A Complete Strategic & Technical Engineering Specification for TrustScan AI.

🌐 PART 1: Non-Technical & Business Aspects

                                  ┌─────────────────────────────────────────┐
                                  │      THE INDIAN FRAUD EPIDEMIC          │
                                  │ 1. Fake Internship & Job Offer Letters  │
                                  │ 2. Forged Aadhaar / PAN Identity Cards  │
                                  │ 3. Fake UPI Transaction Screenshots     │
                                  │ 4. Unregistered MCA Shell Companies     │
                                  └────────────────────┬────────────────────┘
                                                       │
                                                       ▼
                                  ┌─────────────────────────────────────────┐
                                  │             TRUSTSCAN AI                │
                                  │   Multi-Modal Verification Engine       │
                                  └────────────────────┬────────────────────┘
                                                       │
                 ┌──────────────────┬──────────────────┴──────────────────┬──────────────────┐
                 ▼                  ▼                                     ▼                  ▼
        ┌──────────────────┐┌──────────────────┐                 ┌──────────────────┐┌──────────────────┐
        │ 🏢 COMPANY & CIN ││ 🏛️ GOVT ID AUDIT │                 │ 💼 CAREER DOCS   ││ 💳 UPI & BANKING │
        │ MCA Master Data, ││ Aadhaar Verhoeff,│                 │ CTC Math Check,  ││ 12-Digit UTR,    │
        │ GSTIN State Map  ││ PAN Entity Type  │                 │ HR Domain Check  ││ Fake APK Splicing│
        └──────────────────┘└──────────────────┘                 └──────────────────┘└──────────────────┘

1. The Core Problem in India

India faces a digital forgery and financial fraud epidemic:

  • Fake Job & Internship Scams: Unregistered shell agencies send forged offer letters demanding "training fees" or "laptop security deposits".
  • Identity Forgery: Altered Aadhaar cards and PAN cards used for unauthorized SIM cards, loan scams, and KYC fraud.
  • Fake UPI Payment Generators: Android APKs that generate pixel-perfect fake Google Pay, PhonePe, and Paytm success screens to dupe merchants.
  • Corporate Impersonation: Legitimate brand names misused by unregistered entities without MCA (Ministry of Corporate Affairs) or GSTIN registrations.

2. Live Production Telemetry & Real User Demand

Analysis of 19,339 live scans from real users on trustscanai.in:

  • 💬 Text / WhatsApp Messages: 11,790 scans (61.0%)
  • 💳 Payment & Bank Receipts: 7,261 scans (37.5%) — Massive organic merchant demand!
  • 📄 Offer Letters & Documents: 154 scans (0.8%)
  • 📧 Emails & URLs: 129 scans (0.7%)
  • 🏢 Company Verification: 5 scans (0.0%)

3. The 4 Specialized Verification Portals

Instead of a generic scanner, TrustScan provides 4 dedicated verification gateways:

  1. 🏢 Company & CIN Verification: Real-time MCA registry lookup & GSTIN verification.
  2. 🏛️ Government ID Verifier: Cryptographic checksum audit for Aadhaar (Verhoeff $D_5$) and PAN syntax.
  3. 💼 Offer Letters & Credentials: CTC arithmetic balance, recruiter domain audit, and tampering traces.
  4. 💳 UPI & Payment Receipts: 12-digit UTR validation, IFSC branch resolver, and fake APK font analysis.

4. Privacy & Regulatory Compliance

  • Zero PII Retention: Aadhaar, PAN, and banking documents are processed strictly in-memory and wiped immediately after inference.
  • UIDAI Compliance: All displayed Aadhaar references are masked (XXXX XXXX 1234).

⚙️ PART 2: Technical & Engineering Architecture

flowchart TD
    A[User Document Upload: PDF / Image / Text] --> B[Stage 1: Ingestion & Adaptive PyMuPDF Preprocessing]
    B --> C[Stage 2: 23-Language Indic OCR via Sarvam Vision 3B]
    C --> D[Stage 3: Deterministic Mathematical Rules & Checksums]
    B --> E[Stage 3.5: OpenCV / PIL Deep Image Forensics - ELA & Noise]
    C --> F[Stage 4: 11-Feature Document ML Classifier]
    D --> G[Stage 5: Risk-Tiered LLM Reasoning - Sarvam 30B / 105B]
    E --> G
    F --> G
    G --> H[Stage 6: Score Fusion & Deterministic Override]
    H --> I[Stage 6.5: Authenticated Ground-Truth Verification Bridge]
    I --> J[Stage 7: Domain-Specific Interactive Result Dashboards]
Loading

⚡ The 7-Stage Multi-Modal Pipeline Explained

🔹 Stage 1: Ingestion & PyMuPDF Preprocessing

  • File: server/services/processing/documentPipeline.js
  • Ingests PDFs, PNGs, and JPEGs.
  • Converts multi-page PDFs to high-resolution raster images at 2.0x scale via PyMuPDF (fitz) for crisp character edge detection.

🔹 Stage 2: OCR & Layout Digitization (Sarvam Vision 3B)

  • File: server/services/analysis/sarvamService.js
  • Uses Sarvam Vision (3B parameter VLM) fine-tuned on 23 Indian languages.
  • Extracts structured Markdown tables, Devanagari Hindi, Tamil, Telugu, and English text with document hierarchy.

🔹 Stage 3: Deterministic Mathematical Checksums (Zero Hallucination)

Before calling any AI models, deterministic mathematical rules validate the document:

  1. Aadhaar Verhoeff Checksum: Uses the Dihedral Group $D_5$ non-commutative multiplication ($d$) and permutation ($p$) tables: $$c = \sum_{i=0}^{n-1} p(i \bmod 8, d_i) = 0$$ Catches any single-digit replacement or adjacent transposition.

  2. PAN 10-Character Structural Syntax:

    • Characters 1–3: Alphabetic series (AAA to ZZZ)
    • Character 4 (Entity Type): P (Individual), C (Company), H (HUF), F (Firm), A (AOP), T (Trust)
    • Character 5: First letter of Holder's Last Name
    • Characters 6–9: Sequential 4-digit number (0001 to 9999)
    • Character 10: Alphabetic check digit

  3. GSTIN 15-Digit Tax Code:

    • Characters 1–2: State Code (27 = Maharashtra, 07 = Delhi, 29 = Karnataka)
    • Characters 3–12: 10-digit PAN of the entity
    • Character 13: Entity number of the same PAN holder
    • Character 14: Default Z
    • Character 15: Modulo 36 checksum character

  4. MCA 21-Digit CIN Structure: $$\underbrace{\text{L}}{\text{Listing Status}} \underbrace{\text{72200}}{\text{Industry Code}} \underbrace{\text{MH}}{\text{State}} \underbrace{\text{2020}}{\text{Year}} \underbrace{\text{PTC}}{\text{Class}} \underbrace{\text{123456}}{\text{Registration}}$$

  5. CTC & Salary Arithmetic Balance: $$\text{Gross Salary} = \text{Basic} + \text{HRA} + \text{Special Allowances} - \text{Deductions (PF + PT)}$$

🔹 Stage 3.5: Deep Image Forensics Engine

  • File: server/scripts/image_forensics.py + imageForensicsService.js
  • Error Level Analysis (ELA): Recompresses the image at 90% JPEG quality and calculates the absolute difference: $$\Delta_{\text{ELA}} = |I_{\text{original}} - I_{\text{recompressed}}| \times 10$$ Tampered numbers or spliced text exhibit significantly higher compression error variance.
  • Noise Inconsistency: Computes local Laplacian variance across $32 \times 32$ patches.
  • EXIF Metadata Signature Scan: Scans file streams for software tampering signatures (Photoshop, Canva, GIMP, Acrobat).

🔹 Stage 4: 11-Feature ML Document Classifier

  • File: server/scripts/train_document_rules.py
  • Trained Logistic Regression model evaluating 11 document signals: [hasCin, hasGst, mathBalanceValid, officialDomain, softwareSignatures, highUrgencyVelocity, registrationFee, genericTemplate, tamperScore, validPan, verhoeffValid]
  • Achieves 100% fraud recall on adversarial benchmark datasets.

🔹 Stage 5: Risk-Tiered LLM Routing (Sarvam AI)

  • File: server/services/analysis/sarvamService.js
  • Dynamic inference routing:
    • 🟢 Low Risk ($p < 0.3$) $\rightarrow$ sarvam-30b (Fast, cost-effective structured entity extraction).
    • 🔴 High Risk ($p \ge 0.3$ or Checksum Failure) $\rightarrow$ sarvam-105b (Deep multi-step forensic reasoning).

🔹 Stage 6: Multi-Modal Score Fusion & Deterministic Override

$$\text{TrustScore} = \begin{cases} 0% \text{ to } 10% \text{ (FORCED FRAUD VERDICT)}, & \text{if any Verhoeff or PAN checksum fails} \\ (0.40 \times \text{ML_Score} + 0.60 \times \text{LLM_Score}) \times 100, & \text{otherwise} \end{cases}$$

🔹 Stage 6.5: Authenticated Verification Bridge

  • File: server/services/verification/verificationBridge.js
  • Connects to ground-truth public databases:
    • MCA API: Official Ministry of Corporate Affairs Master Data.
    • GST Portal: Active taxpayer status.
    • Razorpay IFSC API: Resolves Bank Name, Branch, City, and State.

🔹 Stage 7: Feature-Specific Interactive Result Dashboards

  • Files:
    • GovIdVerificationCard.tsx — Aadhaar/PAN Checksum, Roboflow Landmarks & ELA Card.
    • PaymentReceiptCard.tsx — UPI UTR Ref & Fake APK Splicing Card.
    • CareerDocumentCard.tsx — CTC Math & MCA Registration Card.
    • BusinessVerificationCard.tsx — MCA 21-digit CIN & GSTIN Registry Card.

🗄️ Database Architecture

Database Technology Purpose
Document Store MongoDB Atlas Stores unstructured scan payloads, multi-lingual OCR extractions, and ELA forensic heatmaps.
Relational Data PostgreSQL User accounts, authentication, scan audit history, and billing records.

🚀 Performance Benchmarks & Scale

  • ⚡ Processing Latency: Reduced from 90 seconds down to <15 seconds (85% optimization via asynchronous Python worker pool).
  • 🌐 Google Search Console Ranking: Average Position #5.1 on Page 1 of Google (8,260+ impressions, 617 organic clicks, 7.5% CTR).
  • 📦 Next.js Production Build: Clean Turbopack builds across 25 static & dynamic routes in ~1.2s.

© 2026 TrustScan AI. All Rights Reserved. Engineered with ❤️ by Shubham Dubey.