AI/ML Engineer — Self-Hosted Document & Order Processing Automation

AI/ML Engineer — Self-Hosted Document & Order Processing Automation We are looking for a strong AI/ML + Python engineer to help build a production-oriented document/order processing automation system. This is not a simple API integration or chatbot project. The system will process business emails and documents, extract structured information, validate it against business rules, route uncertain cases for human review, and generate structured output for an existing ERP workflow. The production architecture is expected to be self-hosted/on-premise, with client-provided GPU/server infrastructure. We do not want the core production workflow dependent on third-party LLM API keys. Core Phase 1 Outcome Build a working pipeline: Email / Documents → Document parsing / OCR → Local AI/ML extraction → Structured JSON/schema → Business-rule validation → Human review for discrepancies → Approved structured data → ERP-compatible CSV/template The system must not silently guess critical fields. Missing, conflicting, invalid, or low-confidence information must be flagged for human review. Expected Technical Scope The engineer will be expected to contribute across the full Phase 1 technical implementation, including email/document ingestion, PDF/OCR processing, structured field extraction, local model inference, validation rules, backend APIs, database integration, background processing, review workflow, CSV generation, testing, deployment, logging, and technical documentation. Likely stack: Python / FastAPI PyTorch / Hugging Face Open-weight LLM/VLM models vLLM or equivalent local inference server Document AI / OCR tooling such as Docling, RapidOCR, Tesseract or suitable alternatives PostgreSQL Redis / Celery or equivalent background processing Docker / Docker Compose Linux / NVIDIA CUDA Frontend may use Next.js/TypeScript or another suitable modern framework. The exact AI model is not predetermined. We expect the engineer to benchmark suitable open-source models against actual sample documents and recommend the appropriate model based on extraction accuracy, hallucination rate, latency and hardware requirements. Custom fine-tuning may be considered where justified by the data, but should not be assumed as the first solution. Important Engagement Terms This is a white-label engagement. You will work as part of our technical delivery team and may occasionally participate in meetings with our end client. During client meetings, your role will be strictly technical: architecture, implementation, debugging, requirements clarification, infrastructure, AI/ML, deployment and related engineering discussions. Commercial discussions, pricing, quotations, contracts, project commercials, future commercial opportunities and business negotiations will be handled only by our team. You must not independently discuss commercial terms, propose separate work, solicit the end client, or attempt to establish a direct commercial engagement with the end client. Confidentiality is mandatory. Project information, source data, architecture, client information, credentials and internal documents must not be shared externally or used publicly without written permission. The project/client may not be used in your portfolio, case studies, social media, proposals or marketing material without written approval. An NDA/confidentiality agreement may be required. Subcontracting or sharing project access/data with another developer is not permitted without prior approval. Ownership & Handover This is not a black-box development engagement. All project-related source code, configuration, prompts, schemas, model configuration, Docker files, database migrations, deployment scripts and technical documentation developed under the engagement must be committed to the designated project repository and handed over as part of the engagement. The solution must be reproducible and deployable by another competent engineer. Avoid unnecessary vendor lock-in or dependencies that prevent us from maintaining the system after handover. Working Style We need someone who is comfortable working on a real production system where requirements will become clearer as actual historical documents and edge cases are tested. The scope described above defines the intended Phase 1 outcome. The individual task list is indicative rather than exhaustive. Normal engineering work reasonably necessary to achieve the agreed Phase 1 outcome—including debugging, integration adjustments, schema changes, deployment configuration, testing and reasonable technical refinements—is expected to be part of the engagement. Materially new business modules or major functionality outside the Phase 1 outcome will be treated separately. We do not want a rigid “that exact line wasn’t mentioned in the original brief” working style for normal implementation details. At the same time, we will not expect unrelated major features to be added without discussing scope. Quality Expectations The system should be production-oriented rather than a demo. We expect: Reliable structured outputs Clear error handling Human review for uncertain values Auditability of extracted/corrected information Secure handling of business data Clean and maintainable code Git-based development Dockerized deployment Reasonable automated/unit/integration testing Logging and debugging capability Deployment documentation Knowledge transfer/handover Support during UAT and resolution of bugs discovered in agreed Phase 1 functionality Ideal Experience Strong Python/backend engineering is required. Experience with local LLM/VLM deployment, Hugging Face/PyTorch, NVIDIA GPUs/CUDA, document AI/OCR, structured LLM extraction, PostgreSQL, Docker and production ML inference is highly preferred. Experience with ERP workflows, SQL Server, email integrations, manufacturing systems, document-heavy automation or LoRA/QLoRA fine-tuning is a plus. We are looking for an engineer who can make technical decisions and solve problems—not someone who only follows predefined prompts or API tutorials. When Applying Please include: 1. Relevant document AI / OCR / local LLM projects you have built. 2. Experience deploying models on local NVIDIA GPU infrastructure. 3. Your expected fixed-price or milestone-based budget range. Please begin your proposal with LOCAL-AI so we know you have read the complete brief.