Smart Invoice Reader

• 

  Live analytics dashboard showing total invoices, spending by vendor, and monthly trends — powered by Gemini AI extraction

• 

  Gemini 1.5 Flash automatically extracts vendor name, invoice number, date, items and amounts from uploaded invoice photos

• 

  Google Cloud Run deployment — containerized FastAPI backend, auto-scaling and always live

• 

  Searchable invoice database with payment status tracking, export to Excel, and per-user data isolation via Supabase RLS

Inspiration

Small businesses in Uganda spend hours manually entering invoice data. We wanted to solve this with AI — making invoice management fast, accurate, and affordable for African businesses.

What it does

Smart Invoice Reader uses AI to automatically extract data from invoice images and PDFs. Users upload an invoice, and our AI engine reads the vendor name, amounts, dates, and line items — then saves everything to a searchable dashboard with analytics.

Key features:

• AI extraction using Featherless AI (Qwen2.5-VL-72B-Instruct) with Gemini as backup
• Multi-user system with secure data isolation (Supabase RLS)
• Live dashboard with charts and spending analytics
• Excel export for accounting
• Duplicate invoice detection
• Free, Starter, and Business pricing tiers (UGX)

How we built it

• Primary AI: Featherless AI — Qwen/Qwen2.5-VL-72B-Instruct vision model for invoice extraction
• Fallback AI: Google Gemini 1.5 Flash as secondary extraction option
• Backend: FastAPI (Python) with async processing
• Database & Auth: Supabase (PostgreSQL with Row-Level Security)
• Deployment: Railway (live production app)
• Frontend: HTML/CSS/JavaScript with Jinja2 templates
• Version Control: GitHub

Why Featherless?

Featherless gave us access to powerful open-source vision models like Qwen2.5-VL-72B-Instruct through a simple OpenAI-compatible API. This model proved highly accurate at reading both printed and handwritten Ugandan invoices — including URA e-receipts, market invoices, and supplier receipts. The reliability and speed of Featherless made it our primary extraction engine.

Challenges we ran into

• Integrating two AI providers (Featherless + Gemini) with automatic fallback
• Implementing strict per-user data isolation with Supabase RLS policies
• Handling both PDF and image invoice formats with PyMuPDF

What we learned

• Featherless AI Qwen2.5-VL-72B-Instruct is exceptionally accurate for reading handwritten and printed invoices in Ugandan contexts
• Building a dual-AI system with automatic fallback increases reliability
• Real African business workflows require UGX currency support

What's next

• Mobile app for on-the-go invoice scanning
• More Featherless models for comparison and accuracy testing
• Multi-currency support across East Africa

Built With

• css
• fastapi
• featherless
• google-cloud-run
• google-gemini-1.5-flash-api
• html
• javascript
• postgresql
• pymupdf
• python
• supabase