🧾 DocuMind-AI — Invoice & Receipt Extraction (Document Intelligence)

🚧 Status: In Progress (under active development)

Full-stack project for extracting key business fields from invoices and receipts (PDF/JPG/PNG), with a built-in review & correction workflow and JSON/CSV export.

📌 Scope (MVP): vendor_name, invoice_date, total_amount, currency, tax_amount
🗂️ Statuses: queued → processing → done / failed
📝 Correction: inline editing + saved as corrected_result

---

---

🎯 Goal

Build an end-to-end system that:

• uploads documents (PDF/JPG/PNG),
• runs an extraction pipeline (OCR / ML),
• stores results and confidence,
• allows manual corrections in UI,
• exports final structured data to JSON and CSV.

---

🧩 What’s inside

├─ backend/ ASP.NET Core Web API + background worker
├─ frontend/ Angular UI (document list, details, corrections)
├─ ml-service/ Python FastAPI inference (stub → OCR → ML model)
├─ docs/ notes, diagrams, roadmap
├─ data/ datasets (ignored by git)
└─ docker-compose.yml

---

🔁 Processing flow (MVP)

1. Upload document → saved to storage + DB record created (queued)
2. Analyze → background worker picks queued docs
3. Worker sets processing → calls ml-service /infer
4. Store raw_result + confidence summary → status done / failed
5. User can correct fields → saved as corrected_result
6. Export final result (corrected if exists, else raw) → JSON / CSV

---

✅ Current progress

This section will be updated as MVP features land.

Backend (ASP.NET Core)

• Basic document processing foundations (statuses + DTOs)
• Storage-related changes (local MVP storage)
• Processing pipeline groundwork (worker/pipeline merged)
• Results storage (document_results: raw/corrected JSON, model_version, confidence)
• Analyze endpoint wiring (POST /api/documents/{id}/analyze)
• Export endpoints (JSON/CSV)

Frontend (Angular)

• Upload page (drag & drop + progress)
• Documents list + filters + pagination
• Document details + extracted fields table
• Inline correction + “edited” highlighting
• Export buttons (JSON/CSV)

ML Service (FastAPI)

• /infer stub (to unblock end-to-end demo)
• OCR baseline switched to paddle
• ML model integration (layout-aware document understanding)

---

🧠 AI / Extraction plan

Phase 1 — Stub (end-to-end demo)

• /infer returns fixed JSON in the final contract format
• Goal: validate pipeline + UI + DB before real extraction

Phase 2 — OCR baseline (practical results fast)

• PDF → images
• OCR with Polish characters support (UTF-8 end-to-end)
• heuristics for totals / tax / dates / currency

Phase 3 — ML model (better generalization)

Train / fine-tune on public datasets (then extend with Polish samples):

• SROIE (receipts)
• CORD (receipts)
• FATURA (Zenodo) (synthetic invoices)

---

📦 Output format (API contract)

{
  "fields": {
    "vendor_name":   {"value":"...",       "confidence":0.82},
    "invoice_date":  {"value":"YYYY-MM-DD","confidence":0.77},
    "total_amount":  {"value":123.45,      "confidence":0.91},
    "currency":      {"value":"PLN",       "confidence":0.88},
    "tax_amount":    {"value":23.45,       "confidence":0.69}
  },
  "model_version": "stub-v0"
}

---

🛠️ Tech stack

Backend: ASP.NET Core
Frontend: Angular
ML Service: Python + FastAPI
Database: PostgreSQL (dev)
Containers: Docker Compose

---

🧑‍💻 Author

Created by Avuii