AI / LLM Engineer

You will build LLM features that run inside customers' ERP workflows: retrieval over their own documents, agents that draft and route transactions, extraction pipelines that replace manual data entry. The bar is production reliability — every feature ships with an evaluation set and a defined failure mode. About the teamThe AI & Data practice is the reason clients pick us over a conventional Odoo partner. We work alongside delivery pods rather than as an isolated R&D group, which means every model you ship has a named user inside a business process on day one. What you'll do Design and ship RAG pipelines over customer documents — chunking, embedding, retrieval strategy, reranking, grounded citation. Build agentic workflows with LangGraph that read from and write to ERP systems safely, with human-in-the-loop checkpoints where the blast radius warrants it. Implement multi-LLM routing: pick the right model per task, control cost, and fail over cleanly. Build evaluation harnesses. A feature is not done until we can measure regressions on a fixed set. Handle the unglamorous 80%: document parsing, OCR quality, schema mapping, idempotency, retries and audit trails. Work directly with consultants and clients to define what "good enough to trust" means for each use case. What we're looking for 4+ years of professional Python, with at least 1–2 years building LLM-backed systems that reached production. Hands-on LangGraph or LangChain experience — you know where the abstractions help and where they get in the way. Practical RAG experience: you've debugged bad retrieval, not just wired up a vector store. Comfortable with at least one vector database (pgvector, Qdrant, Weaviate, Pinecone) and honest about the trade-offs. You design for evaluation and observability from the start, and can explain a failure to a non-technical stakeholder. Overlap with either India, EU or US-East business hours. Nice to have ERP or accounting domain knowledge — you understand why a wrong journal entry is not the same as a wrong chatbot reply. Fine-tuning or preference-optimisation experience. Document AI: OCR pipelines, table extraction, invoice/PO parsing at volume. Experience with self-hosted open-weight models and the cost maths behind that decision. Hiring process Screening (45 min) — a senior engineer on what you've shipped and what broke. Technical deep dive (90 min) — system design for a real RAG/agent problem, plus evaluation strategy. Founder conversation (45 min) — values, ownership, judgement under ambiguity. Decision within 72 hours of the final round.