DocsGuard / Pramaan

A document-authenticity tool combining metadata analysis, image forensics, QR decoding, checksum validation and Claude-assisted reasoning to flag suspicious documents and explain the supporting evidence.

We ran an import/export business. Every deal starts the same way: the counterparty sends their documents — company registration, bank details, certificates, guarantees. We received them, and we had no real way to verify any of it. A single export deal takes 8–12 months, and for that entire time we hold stock reserved for that buyer. We were scammed more than once, and the loss was never just that deal — it was the year, and the inventory we'd held for someone who was never real.

That same gap runs far wider than trade. Forged caste, income and domicile certificates claim welfare benefits (Ladli Behna, scholarships, subsidies, PDS); fabricated marksheets win government jobs; fake turnover and bank-guarantee papers win public tenders. A ₹202 crore ED case involved forged PNB/Bank of Baroda guarantees submitted to Madhya Pradesh Jal Nigam Maryadit; DGGI detected ₹36,374 crore of fake-invoice GST-ITC fraud in FY24–25.

Who experiences it: the clerk at the counter with hundreds of applications and no forensic tooling, the exporter holding stock for a buyer who may not exist, the bank officer accepting a guarantee, and — least visible — the honest applicant rejected because their genuine document looked odd. Both failure directions are invisible today, because nobody records them.

DocsGuard — AI document authenticity & fraud detection

Fraudsters claim public money using documents that were never real: forged income, caste and domicile certificates, tampered marksheets, fake bank guarantees in tenders. Aadhaar proves who you are — it does not prove what your paper says. Today one clerk checks that by eye, across hundreds of applications, with no forensic tooling. A competent forgery passes; an honest applicant with an odd-looking document gets rejected. Both failures are invisible.

DocsGuard turns that check into seconds. Upload any certificate, invoice, ID or statement and it returns a defensible verdict — Authentic, Suspicious or Likely Fake — with a 0–100 risk score.

Core functionality:

Deterministic forensics computed in code — SHA-256 fingerprint, PDF producer and creation-vs-modification dates, image EXIF editor tags, an Error-Level-Analysis heatmap, QR decoding, and checksum validation of Aadhaar (Verhoeff), PAN, GSTIN, IFSC, IBAN and ICAO-9303 MRZ numbers (60/60 tests passing). Claude Opus reads the document — typography and baseline mismatches, copy-paste artifacts, OCR text, and cross-field logic (does the declared income match the bank statement? does the arithmetic hold?). An adversarial review — a prosecution case and a defence case are argued from identical evidence, then adjudicated. The ruling records what was decisive and what it dismissed, which is what prevents false accusations. A 12-module forensic dossier with the evidence quoted, plus an explicit list of what it could not verify.

How it solves the problem: the maths is binding. A number failing its official checksum could not have been issued, so no AI argument can talk the verdict back down — and anything needing a live source (DigiLocker, registry, bank) is flagged for a human rather than invented. That makes every verdict explainable, and an unexplainable rejection is one that gets challenged.

Measured on the deployed app: a tampered income certificate returns LIKELY_FAKE, risk 98, confidence 95, in 40 seconds — with 12 forensic modules, 5 red flags and 21 extracted fields. The adversarial review dismissed 4 prosecution arguments as explainable, which is exactly how honest applicants are protected. Verify the deterministic layer yourself: node --test api/verification.test.mjs → 60/60 passing.