OCR Engine for PDF to Word Conversion

We are looking for an experienced Document AI / OCR / Computer Vision Developer to build a reliable scanned PDF to editable Microsoft Word (.DOCX) conversion engine. This is a focused development project. We already have an existing application. You are not being hired to redesign the application, build the UI, work on mobile development, or develop unrelated features. Your responsibility is only to build the conversion engine: Scanned PDF → OCR → Layout Reconstruction → Editable DOCX Critical Requirement: No Paid OCR or Conversion Services The solution must be developed using our own codebase and open-source libraries/models only. We specifically do not want to use paid OCR APIs, paid PDF-to-Word APIs, commercial conversion SDKs, or third-party SaaS services that charge per page, per document, per API request, or by subscription. The completed system must: Run using our own code Be deployable on our own server/environment Have no per-page conversion cost Have no per-document conversion cost Have no mandatory third-party subscription Have no vendor lock-in Continue working independently of any commercial OCR/conversion provider Include the complete source code for the conversion pipeline Open-source libraries and models are acceptable. Examples may include: PaddleOCR Tesseract OpenCV PyMuPDF pdfplumber python-docx lxml ONNX PyTorch Open-source document layout models Other suitable open-source technologies You may propose another technical stack if it meets the requirements above. Not Acceptable The solution must NOT depend on services such as: Adobe PDF Services ABBYY Cloud Azure Document Intelligence Google Document AI Google Vision API AWS Textract CloudConvert ConvertAPI Aspose paid products Commercial PDF-to-Word APIs Paid OCR SDKs Online PDF conversion websites Any SaaS requiring ongoing payment for conversion We are not hiring someone to simply connect our application to an existing conversion service. We want a reusable conversion engine that we own and control. The Problem We Need Solved We need to convert scanned or image-based PDF documents into Microsoft Word documents while preserving the original document structure as closely as reasonably possible. Basic OCR is not sufficient. We do not simply want extracted text placed into a Word document. The output should resemble the source PDF while still being properly editable in Microsoft Word. Documents may contain: Multiple pages Paragraphs Headings Different font sizes Bold or italic text Indentation Alignment Numbered lists Bulleted lists Tables Headers and footers Page numbers Signatures Stamps Logos Images Mixed text and graphical content Some documents may be legal, business, administrative, or government documents. Required Conversion Behavior The converter must: Accept scanned or image-based PDFs Support multi-page documents Perform OCR automatically Detect page structure and text regions Detect paragraph boundaries Detect document layout Reconstruct the document into DOCX Produce actual editable Word text Preserve reading order Preserve paragraph structure Preserve headings where detectable Preserve alignment Preserve indentation Preserve paragraph spacing Preserve page breaks Preserve bold/italic formatting where detectable Preserve lists and numbering where possible Reconstruct tables into editable Word tables where possible Preserve images, logos, signatures, stamps, and similar graphical elements Preserve headers and footers where possible Preserve general page layout and spacing Avoid duplicated text Avoid missing text Avoid unnecessary floating text boxes Generate a valid DOCX that opens normally in Microsoft Word What We Do NOT Want The following will not be considered a successful solution: 1. Images inserted into Word The system must not simply convert every PDF page into an image and insert those images into Word. Text must be genuinely editable. 2. Plain OCR text dump We do not want OCR text extracted and pasted into Word without attempting to reconstruct the document layout. 3. Manual document recreation The developer must not manually recreate or manually correct the Word document as part of the normal conversion workflow. The conversion must happen automatically through code. 4. Sample-specific hardcoding The system must not be designed only to work with the sample PDF we provide. It must generalize reasonably well across other scanned documents. Expected Architecture We expect the solution to include some form of: PDF/image preprocessing Image cleanup where necessary OCR Text-region detection Document layout analysis Reading-order reconstruction Paragraph reconstruction Table detection/reconstruction Image/signature/stamp handling DOCX generation Error handling Logging or debugging information We are open to better technical approaches proposed by the developer. Measure of Success Success will be measured primarily by the quality of the automatically generated Word document. The project is not considered complete merely because text can be extracted from the PDF. Target Quality For reasonably clear scans, we expect approximately: 98% or better OCR text accuracy We understand that no OCR engine can guarantee perfect accuracy on damaged, blurry, handwritten, or poor-quality source documents. Applicants who promise 100% OCR accuracy on every document will not be considered credible. Acceptance Criteria The conversion engine must pass the following requirements: Text All major text content is present No entire paragraphs or sections are missing Text is editable Text is not embedded as a page image Reading order is correct OCR errors are within an acceptable range Pages All input pages are processed No pages are skipped Page order is correct Major page breaks are preserved Paragraphs Paragraphs remain logically grouped Individual lines should not become unrelated text blocks Paragraph spacing should reasonably resemble the source Layout General visual structure should resemble the source Alignment should be preserved where practical Indentation should be preserved Section placement should remain logical Tables Tables should be detected Table structure should be reconstructed where reasonably possible Tables should preferably be editable Word tables rather than screenshots Images and Other Elements Images, signatures, stamps, seals, and logos should be retained when relevant. They may remain as images where appropriate. Microsoft Word Compatibility The generated DOCX must: Open normally in Microsoft Word Not generate file corruption warnings Not require Word to repair the file Be editable using normal Microsoft Word functions Automatic Failure Conditions The project will automatically fail acceptance if: Any page is completely missing Large sections of readable text are missing The resulting DOCX consists primarily of screenshots Text cannot be edited normally Conversion requires manual corrections by the developer A paid OCR or conversion API is required A commercial SDK subscription is required Documents must be sent to an external paid conversion service The DOCX is corrupted Microsoft Word displays repair warnings The converter only works on the specific development sample Source code is incomplete or unavailable Quality Scoring We will evaluate the completed system approximately using the following weighting: OCR / Text Accuracy — 30% Accuracy of recognized text compared to the source. Layout Fidelity — 25% How closely the generated Word document reflects the structure and positioning of the original document. Word Editability and Structure — 20% Whether paragraphs, tables, headings, and text remain properly editable. Tables, Images, Signatures, and Other Elements — 10% How well non-standard document elements are retained or reconstructed. Reliability Across Different Documents — 10% Whether the solution performs consistently on multiple PDFs rather than only the development sample. Speed and Error Handling — 5% Conversion stability, reasonable processing time, and graceful handling of errors. The target is approximately 90% overall acceptance quality, subject to the automatic failure conditions above. Testing Process We will provide sample scanned PDFs during development. One of the primary test documents is approximately 15 pages and has previously proven difficult for PDF-to-Word conversion. The developer must create a script/module that accepts the document and generates the resulting DOCX automatically. For example: input.pdf → conversion engine → output.docx The developer must not manually edit the resulting Word file after conversion in order to pass the test. Final Testing Final acceptance will NOT rely only on the PDF used during development. We will also test the completed system against additional PDFs that the developer has not previously seen. This is intended to confirm that the solution is generalizable and not hardcoded for one specific document. Deliverables The final deliverables must include: Complete source code OCR processing pipeline Layout-analysis implementation DOCX generation implementation Table-handling implementation Image/signature/stamp handling Error handling Dependency files Configuration files Setup instructions Local/server deployment instructions Required model files or instructions for retrieving open-source models Documentation explaining the conversion flow Example command or API call for conversion Successful completion of our acceptance test documents All source code created for this project must be provided to us. The converter must be ready for integration into our existing application. Development Scope You are ONLY responsible for the PDF-to-Word conversion module. You are not required to: redesign our application develop the application's UI build unrelated application features work on app-store submission work on mobile interface design modify unrelated parts of our system Integration support may be required to expose the finished converter through a clean function, CLI command, or API endpoint. Preferred Technical Background We are particularly interested in developers with experience in: OCR Computer vision Document AI Document layout analysis PDF processing Microsoft DOCX generation OpenCV PaddleOCR Tesseract PyTorch ONNX Python XML python-docx Table detection Document reconstruction Experience specifically generating editable Word documents is highly preferred. Budget Fixed-price budget: $1,000–$1,500 We prefer milestone-based payment. Suggested structure: Milestone 1 — Proof of Concept $150–$250 Demonstrate automated conversion of our provided scanned PDF into DOCX using the proposed open-source approach. Milestone 2 — Complete Conversion Engine $600–$800 Build the complete OCR, layout-reconstruction, table-processing, and DOCX-generation pipeline. Milestone 3 — QA, Edge Cases, Documentation, Final Acceptance $250–$450 Pass additional unseen test documents, resolve major conversion issues, deliver documentation, and prepare the code for integration. Final milestone will only be released after the solution passes acceptance testing. Important Application Instruction Please begin your application with: OPEN SOURCE OCR This lets us know that you have read the full job description and understand that paid OCR/conversion services are not permitted. Applications that recommend simply using Adobe, ABBYY Cloud, Azure, Google Document AI, AWS Textract, CloudConvert, or another paid conversion API will not be considered. We are looking for someone who can actually build the document-processing pipeline, not someone who will simply connect us to another company's PDF-to-Word service.