AI KYC Document Verification

Batch KYC pipeline that OCRs government IDs and supporting documents, structures fields with an LLM, normalises messy values, and runs cross-document verification rules.

Problem

Manual KYC checks across Aadhaar/PAN-style IDs, bank statements, and employment letters are slow and inconsistent. OCR noise (O/0, date formats, phone formatting) breaks naive string equality. This project automates extraction and applies a fixed rule set so reviewers see clear PASS/FAIL outcomes per person.

Features

• Multi-provider OCR: Azure Computer Vision or Google Cloud Vision
• LLM structuring via OpenAI GPT-4 or Anthropic Claude
• Optional OpenCV/Pillow preprocessing for weak scans
• Normalisers for dates, phones, addresses, and case folding
• Seven verification rules (name, DOB, address, phone, father’s name, PAN format, Aadhaar format)
• Dataset runner that writes verification_results.json
• CLI flags for provider choice, preprocessing, and report generation
• Unit-style helpers in test_system.py for normalisers and rules

Architecture

flowchart TD
  Images[Document images per person] --> OCR[OCR provider]
  OCR --> LLM[LLM field extraction]
  LLM --> Norm[Normalisation layer]
  Norm --> Rules[7 verification rules]
  Rules --> Out[JSON results + overall status]

Loading

Application code lives under document-verification-system/.

Tech stack

Layer	Choice
Language	Python 3.8+
OCR	Azure Computer Vision / Google Cloud Vision / optional Tesseract
LLM	OpenAI / Anthropic
Imaging	OpenCV, Pillow, pdf2image
Config	python-dotenv

Key engineering decisions

1. OCR then LLM, not LLM-only vision. Dedicated OCR keeps text extraction controllable; the LLM focuses on schema mapping and light error correction.
2. Rules are deterministic. Format checks (PAN/Aadhaar) and normalised equality are code, not another model call — cheaper and auditable.
3. Provider swappable at the CLI. --ocr-provider / --llm-model let you compare cost and quality without rewriting the pipeline.
4. Fail soft on missing fields. Partial documents produce rule-level FAIL/incomplete signals rather than crashing the batch.

Honest note on metrics: older copy claimed “95%+ accuracy” and fixed dollar costs without a published evaluation harness in-repo. Treat quality as dataset- and scan-quality dependent; measure on your own labelled set before quoting numbers.

Getting started

Prerequisites

• Python 3.8+
• OCR credentials (Azure and/or Google)
• LLM API key (OpenAI and/or Anthropic)

Setup

git clone https://github.com/darkNIGHT669/ai-kyc-document-verification.git
cd ai-kyc-document-verification/document-verification-system
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
cp .env.example .env

Environment variable names (see .env.example):

AZURE_VISION_ENDPOINT=
AZURE_VISION_KEY=
OPENAI_API_KEY=
# plus Google / Anthropic vars if you select those providers

Place person folders under dataset/ (sample layout included). Then:

python run.py
# optional:
python run.py --preprocess --generate-report
python test_system.py

Results write to verification_results.json. A Loom walkthrough may be linked from prior assignment materials.

Project structure

ai-kyc-document-verification/
└── document-verification-system/
    ├── document_verification.py   # orchestration
    ├── image_preprocessing.py
    ├── run.py
    ├── test_system.py
    ├── requirements.txt
    ├── .env.example
    └── dataset/                   # per-person document images

Testing / quality

python test_system.py

Covers normalisers, rule helpers, and JSON shape checks. This is not a full pytest suite with coverage gates.

License

No LICENSE file is present in the repository root at the time of this rewrite. Add one explicitly before public redistribution.