- THE PROBLEM I need to automate data extraction from supplier invoices in PDF and image formats. My management system requires EXACT data matching to function properly. Main technical challenge: The software requires perfect string matching: [Code] + [Description] + [Supplier] Even minor variations (spaces, capitalization) cause recognition errors and require manual intervention. - REQUIRED OUTPUT - CUG LINE The system must produce a structured record (CUG Line) for EVERY LINE of each invoice containing: 1. supplier_name 2. supplier_vat_number 3. customer_name 4. customer_vat_number 5. product_code 6. product_description 7. batch_number (if present) 8. expiry_date (if present) 9. quantity_purchased 10. unit_of_measure 11. unit_price 12. total_price 13. discount (if present) 14. tax_code CRITICAL: Each field must be extracted consistently and deterministically. The same product must ALWAYS produce the same product_description. - TECHNICAL REQUIREMENTS Input: Native PDFs Scanned PDFs Photos of invoices (often taken with phones, poor lighting, angled, blurry) Photo preprocessing required: Photos need automatic enhancement before extraction: Perspective correction (deskew, dewarp) Quality enhancement (denoise, sharpen, contrast adjustment) Lighting correction (remove shadows, uniform brightness) Rotation/cropping (auto-detect document boundaries) Resolution optimization for OCR Main challenges: Each supplier uses different structures Batch/expiry dates in variable positions (same line, next line, 2 lines after, or absent) Need deterministic output for same input CUG Line format must be complete for each product line Photos often have poor quality requiring preprocessing - WHAT I'M LOOKING FOR A solution that: Preprocesses photos automatically for optimal readability Extracts ALL required fields for the CUG Line Guarantees consistency in outputs (same product = same description always) Handles optional fields (batch/expiry might not exist) Learns and improves over time Produces structured output ready for API/webhook - BUDGET < $800 USD for initial MVP development Open to reasonable monthly operational costs for cloud services/APIs. - WHAT TO INCLUDE IN YOUR PROPOSAL 1. IMAGE PREPROCESSING APPROACH Explain: How you'll handle poor quality photos (blur, shadows, angles) Which image enhancement techniques you'll implement How you'll ensure preprocessing doesn't lose important text Automatic vs manual quality checks 2. DETAILED EXTRACTION APPROACH Explain EXACTLY: How you'll extract ALL 14 fields of the CUG Line How you'll handle batch/expiry in variable positions How you'll guarantee deterministic output Output format (JSON, CSV, API call?) 3. SPECIFIC TOOLS & TECHNOLOGY Clearly indicate: Image preprocessing libraries/methods OCR/extraction technology stack and WHY How you handle missing/optional fields How you normalize data for consistency Estimated monthly operational costs 4. HANDLING SPECIAL CASES Explain how you manage: Very poor quality photos (almost unreadable) Multi-page invoices Product lines spanning multiple text lines Batch/expiry in different formats Products without codes (description only) Discounts in various formats 5. SCALABILITY PLAN How the system starts simple and grows How accuracy improves over time How it adapts to new formats Performance with increasing volumes - KEY QUESTIONS TO ADDRESS In your proposal, answer: Photo quality: What's your approach when photos are severely degraded? Preprocessing impact: How do you ensure image enhancement doesn't distort text? Data completeness: How do you ensure extraction of ALL 14 fields with different formats? Batch and expiry dates: What's your strategy for finding them in non-standard positions? Normalization: How do you ensure consistent output for similar inputs? Learning mechanism: How does the system improve accuracy over time? - SUCCESS METRICS The system will be evaluated on: Photo enhancement: % of photos successfully made readable Completeness: % of fields correctly extracted Consistency: same product → always same CUG Line Batch/expiry accuracy: % correct when present Missing field handling: proper signaling of absent fields - EXPECTED DELIVERABLES Complete preprocessing → extraction → CUG Line system Automatic photo enhancement module API/webhook for data delivery Logging system to track accuracy Performance monitoring capability CUG Line format documentation Source code with documentation - IMPORTANT NOTES System must handle very poor quality photos automatically CUG Line must be complete even if some fields are empty Deterministic: same document → always same CUG Line Scalable: must handle growing volumes Learning capability: improve over time especially on batch/expiry extraction - SELECTION CRITERIA I will evaluate proposals based on: Image preprocessing strategy and experience Technical clarity and feasibility Understanding of the CUG Line requirement Strategy for difficult fields (batch/expiry) Normalization approach Realistic accuracy estimates over time - TO APPLY Start your proposal with: Your image preprocessing approach for poor quality photos Your proposed technology stack and why it's optimal Specific strategy for variable batch/expiry extraction Normalization method for consistent outputs Realistic accuracy timeline (1 month vs 6 months) Monthly operational cost estimate Generic proposals will be ignored. I want to understand EXACTLY how you'll build this system, especially the photo preprocessing pipeline. Note: The ideal solution should be pragmatic and incremental, not perfect from day one. The system must be able to evolve without complete rewrites.