Japanese OCR / Document AI SDK QA Engineer (Linux & Python)

PROJECT OVERVIEW We are looking for an experienced QA engineer to evaluate a document extraction and OCR SDK for Japanese-language documents. Budget: USD 1,500 fixed price Duration: Approximately 10 business days of active work Target completion: August 31, 2026 Start: As soon as possible The SDK runs on Linux x64. The selected contractor will install the SDK, test it against Japanese invoices, contracts, forms, and scanned documents, measure extraction accuracy, document reproducible defects, and develop an automated regression test program. This project is focused on measuring and documenting the SDK's current quality. Improving the OCR engine to reach a specific accuracy target or making major changes to the SDK itself is not part of the scope. The SDK, license, English documentation, and available test materials will be provided after contractor selection and, if required, execution of an NDA. SCOPE LIMITS - Up to 50 document files, with a maximum of 150 pages in total - One agreed Linux x64 environment - Up to 20 Japanese UI screens and 30 pages of documentation for language review - One initial test cycle and one corrected-SDK retest - One agreed set of extraction fields - Waiting time for a corrected SDK is not included in the 10 active business days SCOPE OF WORK 1. Environment setup - Prepare or use a Linux x64 test environment - Install the SDK and required dependencies - Configure licensing or authentication - Confirm supported input and output formats - Run an initial smoke test - Review SDK errors and logs 2. Test planning - Classify document types and extraction fields - Define test cases and expected results - Define the ground-truth data format - Define accuracy metrics and defect severity levels - Confirm acceptance criteria before full testing 3. Test data and ground truth - Organize up to 50 approved documents / 150 pages total - Create or normalize expected results in JSON or CSV - Map each test document to its expected output - Ensure that personal and confidential information is handled appropriately Expected document types may include: - Japanese invoices, contracts, application forms, and business forms - Scanned, skewed, noisy, or low-resolution documents - Documents containing vertical Japanese text - Mixed kanji, hiragana, katakana, Latin characters, and numbers - Tables and line items - Stamps or seals - Dates, amounts, currencies, names, companies, and addresses - Multi-column or otherwise complex reading order Test data may consist of approved public data, synthetic data, or data supplied securely by the client. 4. OCR and extraction testing Evaluate items such as: - Company and personal names - Addresses, telephone numbers, and email addresses - Invoice, document, and contract numbers - Issue dates and payment due dates - Subtotals, tax, totals, and currencies - Product or service names, quantities, and unit prices - Line items and table structures - Stamps or seals, where supported - Text reading order Also record crashes, timeouts, encoding problems, and unexpected errors. 5. Accuracy evaluation Compare SDK output with ground truth and classify results as: - Exact match - Partial match - Missing extraction - Incorrect extraction - Incorrect field assignment - Character corruption - Broken table structure - Incorrect reading order Where appropriate, calculate field-level accuracy, document-level accuracy, precision, recall, F1 score, and character or word error rate. Reaching a specific accuracy level is not an acceptance requirement. 6. Defect investigation Each reported problem should include: - Issue summary and severity - Affected document and input conditions - Expected and actual results - Reproduction steps - Relevant SDK output and logs - Screenshot or other supporting evidence - Reproduction frequency - Suggested workaround or improvement, where possible 7. Automated regression test program Develop a reproducible test program, preferably in Python, that can: - Process all documents in a specified directory - Execute the SDK automatically - Save raw SDK output - Compare output with JSON or CSV ground truth - Detect differences and determine pass/fail status - Aggregate results by document and field - Export results in CSV and/or JSON - Save execution logs - Compare the current run with a previous run Another language may be used if required by the SDK interface. 8. Japanese UI and documentation review Review the available Japanese UI or Japanese-facing materials for: - Unnatural Japanese - Translation errors or inconsistent terminology - Buttons and error messages - Display problems in a Japanese environment - Missing or unclear setup instructions - Areas likely to confuse Japanese users If editable source files are unavailable, provide proposed corrections in the final report. 9. One regression retest If a corrected SDK is supplied within the agreed schedule, perform one retest to confirm: - Previously reported problems have been addressed - Existing document processing still works - Accuracy has not materially regressed - No new crashes or major errors have appeared DELIVERABLES - Test plan and test case list - Approved test data and structured ground-truth data - Regression test source code - Environment setup and execution instructions - Document-level and field-level accuracy results - Defect list - Reproduction steps, logs, screenshots, and supporting evidence - Japanese UI and documentation improvement proposals - Results of one corrected-SDK retest, if the SDK is supplied on schedule - Final report and handover materials Any restrictions on redistributing third-party or confidential test data must be documented. ACCEPTANCE CRITERIA The project will be accepted when: - The SDK can be executed in the agreed Linux x64 environment - Up to 50 agreed documents / 150 pages have been tested - Results are recorded for the agreed extraction fields - Reported defects contain evidence and reproduction instructions - The regression program can be rerun using the provided instructions - Source code and accuracy reports have been delivered - Japanese UI and documentation issues have been documented - One retest has been completed if the corrected SDK is provided within the agreed schedule - Final reporting and handover are complete Acceptance does not require the SDK to reach a specific accuracy level or for every reported defect to be fixed. OUT OF SCOPE - Major SDK source-code modifications - Development of a new OCR engine - AI model training or fine-tuning - Production integration - Building or operating commercial infrastructure - Testing beyond the agreed document/page limit - More than one corrected-SDK retest - Extensive UI or documentation rewriting outside the agreed limits - Ongoing production-data processing - Guaranteeing OCR accuracy Additional work will require a separate estimate and milestone. SECURITY REQUIREMENTS - Do not upload documents to external services without written approval - Do not reuse test data for another purpose - Protect personal and confidential information - Sign an NDA if required - Follow the agreed deletion procedure after completion - Do not disclose the SDK, source code, or test results to third parties REQUIRED QUALIFICATIONS - Professional or native-level Japanese - Experience testing OCR, document AI, or document extraction systems - Strong Python test-automation experience - Linux x64 development and troubleshooting skills - Experience working with JSON, CSV, APIs, logs, and command-line tools - Ability to build structured ground-truth datasets - Understanding of precision, recall, F1, and OCR error metrics - Clear written reporting in English - Ability to handle confidential materials securely Experience with Japanese invoices, contracts, table extraction, reading-order evaluation, or image preprocessing is a plus. PROPOSED MILESTONES 1. SDK setup, smoke test, test plan, and test-case design: USD 250 2. Ground truth, full testing, accuracy evaluation, defect evidence, and regression program: USD 950 3. One corrected-SDK retest, final report, and handover: USD 300 Total: USD 1,500 PLEASE INCLUDE IN YOUR PROPOSAL 1. Whether you can complete the project 2. Your earliest available start date 3. Whether you can complete approximately 10 business days of active work by August 31, 2026 4. Number of team members and their roles 5. Relevant OCR, document AI, Japanese-language testing, and automation experience 6. Estimated effort for each project phase 7. Proposed technologies and test environment 8. What you require from us before starting 9. Confirmation that you accept the USD 1,500 fixed budget 10. Assumptions, possible additional costs, and current questions Please briefly describe one relevant OCR or document-processing project and your specific contribution.