Senior AI/ML Engineer (LLM/RAG/Agents) + Python Backend

Description We’re building an AI-powered knowledge assistant that lets our team query internal documents (policies, reports, support docs) through natural language and get accurate, sourced answers, plus a set of automated agent workflows that act on that information (triage, summarization, routing). We need a senior engineer who can own the AI/ML layer end-to-end and is also comfortable shipping the backend services around it. This is architecture-level LLM engineering plus production Python, not prompt-tweaking. AI/ML work Design and build a RAG pipeline: chunking strategy, embedding model selection, vector store (open to your recommendation: Pinecone, Weaviate, or pgvector, tell us your pick and why) Build and debug multi-step agent workflows (open to your recommended framework: LangGraph, LlamaIndex, CrewAI, or custom orchestration) Prompt engineering and evaluation. You should be able to talk about eval frameworks (e.g. RAGAS, custom eval harnesses), not just “it works when I test it” Handle hallucination mitigation, context window management, and cost/latency tradeoffs between models (GPT-4-class vs smaller/cheaper models) Fine-tuning or lightweight adaptation experience is a plus (LoRA, embeddings fine-tuning). Not required, but tell us if you’ve done it Backend work Build the FastAPI services that expose the AI pipeline: auth, request handling, streaming responses Design the data layer: Postgres schema, background job handling (Celery/RQ), caching (Redis) Write production-grade code: tests, logging, error handling. Not notebook code shipped as-is Must-Haves 3+ years Python, 1.5+ years hands-on with LLM application development (not just API calls) Real production RAG or agent-system experience, can speak to tradeoffs not just tool names FastAPI or equivalent async Python backend experience Git-based workflow, comfortable with code review Nice-to-Haves Experience with AWS Bedrock, Azure OpenAI, or self-hosted models via vLLM Vector DB tuning experience at scale (millions of documents)