Senior AI Full-Stack Developer for OCR, RAG, Gmail & Google Sheets Integration

We are looking for an experienced AI full-stack developer to build a production-ready web assistant for a private healthcare company in the European Union. This is not a simple ChatGPT chatbot. The system must process Hungarian laboratory referral documents, find the requested laboratory tests and match them with a live price list. Two anonymized sample documents are attached. More anonymized samples will be shared only with shortlisted developers. Main functions 1. Embedded web assistant The assistant must be embedded directly into an existing website. It must not be a popup. Users must be able to upload images and PDF files. The design must work well on desktop and mobile. During processing, the user must see clear status messages, for example: Uploading document Reading the referral Finding laboratory tests Checking current prices Preparing the result 2. Referral document recognition The system must recognize Hungarian laboratory referrals, including: printed documents; handwritten text; checked boxes and X marks; tables with several columns; medical abbreviations; documents with low image quality, rotation or shadows. The system must extract only the requested laboratory tests. It must also understand rules such as: one checked row may contain several tests; different abbreviations may mean the same laboratory test; some items are packages; some words are general categories and need a clinic-specific rule; the diagnosis must not be used to invent additional tests. We will provide: anonymized referral samples; the correct expected result for test documents; Hungarian synonyms and abbreviations; clinic-specific interpretation rules; examples of difficult or uncertain cases. The developer does not need to speak Hungarian, but the system must support Hungarian text correctly. Accuracy requirement Our target is at least 97% item-level recognition accuracy on the agreed test dataset. We will measure separately: missed laboratory tests; incorrectly added tests; correct matching with the price list; correct handling of uncertain documents. The system must not guess when the document is unclear. For low-confidence cases, it must: ask the user to upload a better image; show the recognized items for confirmation; or send the case to human review. A 100% accuracy guarantee is not required, but the error handling must be safe and clear. 3. Google Sheets price list The official price list is stored in Google Sheets. The system must: connect directly to Google Sheets; always use the latest data; not require manual AI retraining after a price change; find tests using official names and alternative names; calculate prices only from the current Sheet data; never invent a price. The price list may also include: preparation information; fasting requirements; result delivery times; notes and alternative test names. 4. Knowledge-based questions Users must also be able to ask normal questions. The assistant may answer only from: approved website content; the Google Sheets data; approved FAQ content; clinic-specific rules; approved reply templates. It must not use information from other clinics or external medical websites. If no approved answer is available, it must clearly say that it cannot give a reliable answer and offer human assistance. 5. Gmail integration The system must connect to Google Workspace Gmail. For incoming emails, it must: read the email; prepare a reply draft; create the draft in the original Gmail conversation; apply a label such as “AI review required”; allow a staff member to review, edit and send the draft. The first version must not send emails automatically. Automatic sending may be a later phase only for safe and approved question types. 6. Modular AI architecture The system should use one AI provider in the first production version. However, the architecture must make it possible to replace the AI provider later without rebuilding the complete application. For example, it should be possible to change between: OpenAI; Google Gemini; Anthropic Claude; another compatible provider. A provider change may require testing, but the main application should stay independent from one AI company. 7. Hosting, ownership and security The system must be self-hostable. Requirements: Docker-based deployment is preferred; full source code must be stored in our GitHub repository; all service accounts must be owned and paid directly by the client; no hidden monthly maintenance dependency; clear setup and deployment documentation; secure temporary file handling; sensitive files must not appear in application logs; automatic deletion of temporary uploaded files; GDPR-conscious architecture; access control and basic audit logging. The client will use a lawyer for the final data protection review. The developer must clearly explain the complete data flow. Pilot phase We want to start with a paid technical pilot before the full development. The pilot should include: a simple document upload interface; referral image processing; extraction of requested laboratory tests; matching with sample Google Sheets data; confidence scoring; handling of uncertain cases; an accuracy report. We will provide development samples and separate unseen test samples. If the pilot is successful, the same technical core should be used in the final system. Please include in your proposal Please answer all questions below: What similar OCR, document AI, RAG, Gmail API or Google Sheets projects have you completed? What technology stack would you use? How would you process handwritten text and checked boxes? How would you measure the 97% accuracy target? How would you prevent unsupported or invented answers? How many hours do you estimate for the pilot? How many hours do you estimate for the complete system? What is your realistic delivery time? What monthly external software and API costs do you estimate for approximately 500 requests? Will you personally complete the work, or will other team members be involved? Can you commit all work regularly to our GitHub repository? Can you provide at least 30 days of bug fixing after final delivery? Please do not send a generic AI-generated proposal.