// the find
tukaramchate/Document-Validator
🛡️ AI-Powered Document Validator | React + Flask + FastAPI + PyTorch | Multi-stage verification with CNN forgery detection, OCR (Gemini + fallback), institutional record cross-checking, and PDF report generation.
VeriAcd is a student project that validates academic documents (marksheets, certificates) using a CNN forgery detector, OCR via Gemini, and cross-checking against institutional records. It's a three-service setup: React frontend, Flask business logic, and a FastAPI AI microservice. The target audience is anyone building or studying document verification pipelines, not anyone deploying this in production.
The architecture separation between inference (FastAPI/PyTorch) and business logic (Flask) is sensible — the AI service can be scaled or swapped independently. Including a Gemini OCR path with a fallback parser shows awareness that cloud OCR will occasionally fail. The training data is committed to the repo, so you can actually see what the CNN was trained on and reproduce the model rather than getting a black-box checkpoint. Flask-Limiter is wired in from the start, which is a better default than most student projects.
The training dataset is almost entirely augmentations of a single document (Upto-4th-Sem-Markscard) — that CNN will likely overfit badly to that one layout and fail on anything structurally different. The 'institutional record cross-checking' is vague; there's no indication of how institutional records are populated or maintained, so it's unclear whether this actually catches forgeries or just checks a local lookup table you seeded yourself. Running three separate Python environments manually via PowerShell scripts is fragile for anything beyond a laptop demo — no Docker, no compose file. The repo is tagged as C in GitHub's language detection, which is almost certainly wrong and suggests the language metadata was never corrected, a small but telling sign of early-stage polish.