💡 Inspiration

Across India, digital fraud has caused devastating financial losses exceeding ₹1,750+ Crores quarterly, particularly impacting non-technical organizations such as local retail shops, schools, community clinics, and everyday citizens. These users face a rising tide of multi-vector social engineering:

Fake UPI Payment Receipts: Scammers use online receipt generators or photo editors to alter PhonePe/Paytm/GPay amounts from ₹150 to ₹1,50,000, deceiving retail merchants.
Malicious QR & Collect Hooks: mode=02 reverse-debit traps trick victims into approving withdrawals instead of receiving money.
Predatory Job & Internship Offers: Fraudulent appointment letters demanding upfront "laptop" or "training" fees targeting students.
Digital Arrest & Extortion Calls: Real-time AI face-swaps and video deepfakes impersonating law enforcement.
Manipulated Invoices & Documents: Altered tax bills, tampered PDFs, and arithmetic discrepancies.

Non-technical personnel lack the budget, staff, and jargon-heavy dashboards of enterprise Security Operations Centers (SOCs). We built SatyaKavach (सत्य कवच) to act as an automated, accessible cyber-forensic shield for Bharat.

🛡️ What SatyaKavach Does

SatyaKavach is an enterprise-grade 5-in-1 digital forensics workbench and security copilot featuring:

UPI Payment Screenshot Inspector: Performs client-side HTML5 Canvas Error Level Analysis (ELA) at 75% recompression to render pixel-tampering heatmaps, algorithmic 12-digit NPCI Julian UTR validation, and bank VPA handle checks (@ybl, @okaxis, @paytm).
Malicious QR & Trap-Link Scanner: Dissects raw upi://pay payloads before execution, immediately flagging reverse-debit collect parameters (mode=02) and checking domain threat intelligence.
Job Offer Letter Screener: Multi-modal OCR and linguistic parsing to intercept fake hiring letters demanding upfront fees or using typosquatted corporate HR domains.
Deepfake Call & Video Inspector: Frame-by-frame facial landmark jitter tracking, blink cadence anomalies, and "Digital Arrest" script pattern flags.
Document & Invoice Forensics: Binary metadata inspection, EXIF editor detection (Adobe Photoshop, Canva traces), and itemized arithmetic verification.
Kavach Copilot (AI Security Agent): An interactive bilingual (English/Hindi) conversational agent providing immediate plain-language triage and automated module routing.
Legal Compliance & Court-Admissible Dossier: Automatically generates SHA-256 sealed PDF/JSON evidence records compliant with Section 63 of Bharatiya Sakshya Adhiniyam (BSA 2023) and Section 65B of the Indian Evidence Act, coupled with a 1-click formal complaint generator for India's National Cyber Crime Helpline (1930 / cybercrime.gov.in).

⚡ How We Built It & NVIDIA / Nebius Integration

SatyaKavach is engineered with a Zero-Retention Hybrid Architecture:

Frontend & Edge Routing: Built with Next.js 14 (App Router), TypeScript, and Tailwind CSS, deployed on the Vercel Edge Network for sub-2-second global latency.
NVIDIA Nemotron & Nebius Token Factory:
- We route queries to NVIDIA Llama-3.1-Nemotron-70B-Instruct via Nebius Token Factory / Serverless Endpoints to power the high-order semantic reasoning of our Kavach Copilot and multi-modal document fraud audits.
- Faster Nemotron variants handle low-latency user intent classification and Hindi/Hinglish language translation, maintaining conversational fluidity while keeping token consumption economical.
Deterministic Algorithmic Fallback: Client-side HTML5 Canvas ELA, regex NPCI Julian UTR parsing, and GSTIN checksum calculators ensure immediate verification even under constrained connectivity.
Privacy by Design: File inputs are processed purely in-memory (V8 ArrayBuffer / Uint8Array) and garbage-collected post-inference. Zero bytes are written to disk, and citizen data is never stored or used in training sets.

🧗 Challenges We Ran Into

Sub-2-Second Multi-Modal Inference: Running deep linguistic reasoning over complex offer letters and documents while preserving quick turnaround required setting up efficient prompt schemas and streaming endpoints via Nebius Token Factory.
Explainable Canvas ELA in the Browser: Developing a lightweight, real-time Error Level Analysis canvas algorithm that renders thermal heatmaps without crashing mobile browsers or client threads.
Dual-Language Accessibility: Making cyber-forensics terms accessible to non-technical Indian merchants in natural Hindi and English without losing forensic accuracy.

🏆 Accomplishments That We're Proud Of

Sub-1.8s End-to-End Latency: Complete forensic pipeline—from image upload to Canvas ELA and Nemotron reasoning—finishes in under 1.8 seconds.
BSA 2023 / Section 65B Compliance: Built real cryptographic SHA-256 time-stamped custody chains suitable for actual legal reporting.
Production-Ready Prototype: Live and interactive at https://satya-kavach-live.vercel.app/ with 1-click test vectors for immediate juror evaluation.

📖 What We Learned

Deploying LLMs via Nebius Token Factory Serverless Endpoints dramatically reduces operational complexity compared to managing self-hosted infrastructure.
Combining deterministic forensic mathematics (ELA, UTR checksums) with LLM semantic reasoning creates a far more resilient fraud-detection barrier than relying on machine learning alone.

🚀 What's Next for SatyaKavach

Digital India Bhashini Voice Integration: Expanding the Kavach Copilot to support 12+ regional Indian voice dialects (Tamil, Telugu, Bengali, Marathi).
WhatsApp & Telegram Gateway: Zero-install chat gateways enabling rural merchants to forward suspicious screenshots directly for instant forensic verification.
B2B SDK & Law Enforcement Webhooks: Direct API integrations for banking Fraud Monitoring Units (FMUs) and state cyber cells.