IDP / Document AI

Formly IDP architecture

IDP architecture for long, unstructured documents: OCR, LLM orchestration, Smart Anchor extraction, async backend, queues, caching, and S3 storage.

Client Work 2025
Python FastAPI asyncio Azure Document Intelligence PyMuPDF Pydantic Docker AWS S3

Estimated form-filling time reduced from 40 minutes to about 3 minutes.

>95% extraction accuracy reported for 20-50 page documents.

Context

Antonii worked as an AI Integration Engineer on IDP architecture for processing long, unstructured documents. The case is private/client work, so implementation details are intentionally summarized.

Problem

Manual form filling from 20-50 page documents is slow and error-prone. The system needed to extract structured data while handling layout variation and imperfect OCR.

Constraints

  • Long unstructured documents.
  • OCR quality and page layout variability.
  • Async backend requirements with queues, caching, and storage.
  • Production-MVP expectations around reliability and reviewability.

Approach

The architecture combined OCR, PyMuPDF processing, Smart Anchor extraction, LLM orchestration, Pydantic validation, FastAPI, queues, caching, Docker, and AWS S3 storage.

Implementation highlights

  • Smart Anchor algorithm for stable extraction from variable document layouts.
  • LLM orchestration around OCR and PDF parsing outputs.
  • Async FastAPI backend with queues and caching.
  • Pydantic models for structured validation.

Result

Reported outcomes included estimated form-filling time reduced from 40 minutes to about 3 minutes and more than 95% extraction accuracy for 20-50 page documents.

What it demonstrates

Document AI architecture, async backend engineering, LLM orchestration, OCR pipeline design, and careful handling of private client work.

01Document upload
02OCR
03Smart Anchor extraction
04LLM orchestration
05Validation
06Async backend
07S3 storage