Automating Invoice Processing with OCR + LLMs + FastAPI + Streamlit

Over the last few days, I built a small but powerful end-to-end invoice automation system — from PDF upload to structured data, with duplicate detection to prevent fraud and re-processing.

The goal was simple:

• eliminate manual data entry
• automatically extract invoice fields
• block duplicate invoices
• expose APIs and UI for business users

What the system does

• Upload invoice PDFs
• Extract text using OCR
• Use LLM to parse key fields (invoice no, vendor, date, total)
• Store invoices in SQLite
• Detect duplicate invoices via checksum
• View stored invoices in Streamlit dashboard
• Built using FastAPI backend APIs

This turned into a very practical AI + backend engineering exercise applicable to AP automation and fintech workflows.

High-Level Architecture

+------------------+
|  PDF Invoice     |
|  Upload (UI)     |
|  Streamlit       |
+--------+---------+
         |
         v
+------------------+
| FastAPI Backend  |
+------------------+
         |
         v
+------------------+       +-------------------+
|  OCR Layer       |       | LLM Extraction    |
| (PDF -> text)    |-----> | (Field Parsing)   |
+------------------+       +-------------------+
         |
         v
+------------------------------+
| Duplicate Detection Module   |
| (checksum + DB lookup)       |
+------------------------------+
         |
         v
+------------------------------+
|  SQLite Database             |
|  invoices table              |
+------------------------------+
         |
         v
+------------------------------+
| Streamlit Viewer             |
| list + search invoices       |
+------------------------------+
        

Key AI Component — LLM Field Extraction

Instead of brittle regex-only parsing, an LLM helps map messy OCR text into structured JSON fields such as:

• invoice_number
• vendor
• invoice_date
• total_amount

This allows support for:

✔ different invoice layouts ✔ different vendors ✔ noisy OCR text

🔐 Duplicate Invoice Detection (very important)

To prevent:

• double payment
• invoice fraud
• resubmissions

I added:

• PDF text checksum
• DB uniqueness constraint

Flow:

1. extract text
2. compute SHA256 hash
3. check if checksum already exists
4. if exists → block storage and flag duplicate
5. else insert into DB

This mirrors real AP automation systems.

What I learned

• OCR + LLMs complement each other
• Databases still matter in AI projects
• Duplicate detection must be designed early
• Small automation = big business value
• FastAPI + Streamlit = super quick prototyping stack

Tech Stack

• FastAPI
• Streamlit
• SQLite + SQLAlchemy
• Python OCR
• LLM prompt parsing
• REST APIs

✅ Real-world applications

• Accounts Payable Automation
• Fintech onboarding
• Expense invoice verification
• Procurement systems
• ERP integrations

https://github.com/nagrgkgc/invoice_ai