Lumint is a free, open-source fraud detection platform for India's digital payment ecosystem. It analyzes UPI screenshots, documents, and suspicious URLs — all without your data ever leaving your device.
Lumint runs the entire analysis pipeline in your browser. No uploads, no accounts, no telemetry.
Drop a UPI screenshot, document, or paste a suspicious URL. Files never leave your device — they stay in browser memory.
Tesseract.js extracts text, Canvas-based ELA surfaces pixel tampering, and a rule engine weighs every signal.
A risk score (0–100) with a plain-English explanation of every signal that contributed. Copy, share, or re-run instantly.
Each module targets a specific fraud surface. Run any one independently or chain them through Fraud DNA.
Verify PhonePe, Google Pay, Paytm, and BHIM payment screenshots.
Reads UTR, amount, sender/receiver VPA, and payment app via Tesseract.js OCR. Cross-checks 4-of-5 signal heuristics (payment keywords, app name, @VPA, UTR pattern, rupee symbol) to reject non-UPI inputs like college IDs or random photos.
Open UPI ShieldDetect tampered Aadhaar, PAN, invoices, and certificates.
Runs Error Level Analysis (ELA) by recompressing pixels in-browser to expose JPEG-block tampering, plus metadata/EXIF inspection and font-baseline checks. Flags both crude edits and subtle redactions.
Open DocShieldIdentify lookalike bank domains and UPI phishing links.
Analyzes URLs for typosquatting (paypa1.com), suspicious TLDs (.tk, .ml, .ga), IP-address hosting, HTTPS misuse, and brand impersonation. Each signal gets its own weighted contribution to the final risk score.
Open PhishShieldVisualize fraud campaign networks and shared infrastructure.
Clusters threat events by shared indicators — domains, file hashes, VPA patterns, text fingerprints — and renders the resulting graph. Useful for spotting the same attacker across multiple reports.
Open Fraud DNALumint is built around a single principle: a fraud tool that ships your data to a server is itself a fraud risk.
OCR, ELA, and the rule engine all run in your browser via WebAssembly + Canvas. The app works fully offline once loaded.
There is no sign-up, no analytics on your scans, and no fingerprinting. Open the page, use it, close the tab.
Every risk score is broken down into the signals that contributed. You can audit the rules in the open-source repo.
Tesseract.js (OCR)
Text extraction runs on a WebAssembly port of Google's Tesseract engine. The OCR model (~2 MB) loads once per session and is cached by the browser. Every character, digit, and symbol comes from your local CPU — no API call.
Canvas + WebGL (ELA)
Error Level Analysis recompresses your image at a known JPEG quality, then subtracts it from the original pixel-by-pixel on an HTML5 canvas. Regions that were edited after the original compression show up as bright anomalies — a classic signal of tampering.
Rule engine
Extracted text and ELA scores feed a weighted rule engine. Each rule contributes additively to the final 0–100 risk score, and the explanation panel shows exactly which rules fired and how much each one weighed.
No backend services. No ML APIs. No vendor lock-in.
Next.js 15
Framework
TypeScript
Language
Tailwind CSS v4
Styling
Tesseract.js
OCR (WASM)
Framer Motion
Animation
Canvas / WebGL
Forensics
Vercel
Hosting
MIT License
License
B.Tech CSE · Class of 2027
Tanmay is a Computer Science undergraduate at VIT Bhopal, where he's been building applied AI products since his first semester. Lumint started as a side project to understand how client-side forensics could replace server-side AI APIs — and grew into a full multimodal fraud intelligence framework.
He previously interned at TradeVed (fintech, 2024) and MAET (Marine & Automotive Engineering Training, 2023), working on data pipelines, internal tooling, and ML-assisted product features.