Follow stream · CASE 05 · Digital Forensics · FinTech Security
VeriFlow
VeriFlow Suraksha
Document DNA for loan underwriting.
Live demo
Deployed demo with five pre-loaded test packets.
An offline, deterministic document-forensics engine for bank underwriting. It checks loan-applicant packets (salary slips, bank statements, employment letters, ITRs) for tampering, inflated figures and chronological anomalies, and every flag comes with its evidence and its arithmetic.
01 Lab: check a loan packet
Change any figure and watch the coherence rules re-evaluate, with the arithmetic shown.
Four of VeriFlow's twelve rules, re-implemented in the browser. Edit any figure. The real engine also checks font geometry, error level analysis, PDF chronology and tax recalculation, and routes internally consistent forgeries (packet E) to manual review.
02 Brief
The problem
Loan fraud rarely looks like fraud. A well-made forgery is internally neat. The tells are in the file structure and in whether the numbers agree across documents. Black-box AI tools give underwriters a score they cannot defend.
Why it matters
Underwriting decisions have to be explainable. A forensic engine should be able to show its work like an accountant, not just output a probability.
03 Architecture
A two-tier forensic pipeline. Layer 1 inspects file structure before reading any text. Layer 2 extracts financial fields and cross-checks them across the whole packet.
- 01
Structural DNA
Font-size micro-deviations (0.1–0.8 pt), baseline drift, character spacing: the marks copy-paste editing leaves in PDFs.
- 02
Error Level Analysis
JPEG recompression inconsistencies that reveal pasted regions in scanned documents.
- 03
Chronological forensics
Suspicious PDF producers, IST timezone mismatches, creation/modification inconsistencies.
- 04
Coherence engine
Cross-document maths (salary slip vs bank credit vs loan application) and independent TDS / PF recalculation.
- 05
12-rule forensic engine
Arithmetic fraud, impossible values, round-number suspicion, CTC vs monthly mismatch and more.
04 Implementation
Stack
Python · FastAPI · PyMuPDF · OpenCV · NumPy · EasyOCR · Pydantic v2 · React 19 · Vite
My contribution
Sole committer on the public repository.
- Runs entirely offline, so applicant documents never leave the machine.
- TDS is recalculated under the FY 2024-25 new-regime slabs, and PF at 12% of basic, to catch fabricated deductions.
- Five pre-loaded packets cover genuine, income mismatch, arithmetic fraud, visual anomaly and a perfect forgery that routes to manual review.
05 Engineering decisions
Deterministic over probabilistic
Every decision is reproducible and backed by exact calculations, because an underwriter has to defend a rejection.
Structure before text
Inspecting how a file was built catches edits that leave the visible text perfectly plausible.
06 Evidence
No published results yet. This file is at the live demo stage, and results will be added once they can be backed by evidence.
Sourced from: github.com/Prathamesh0-0/VeriFlow-Suraksha README; PRD.md