Staff AI Platform Engineer, Afresh

• 5+ years building production software, data, or ML systems; an excellent engineer with strong systems and API design (Python)
• Hands-on production experience with LLM systems: retrieval/RAG, agents and tool-use, prompt and context engineering — and, critically, evaluation. You measure quality; you don't eyeball it
• Solid data-engineering and data-platform foundations: pipelines, data modeling, and a modern cloud data stack (Databricks/Spark, MLflow, cloud warehouses)
• Comfort in the messy middle of AI systems — retrieval quality, latency and cost trade-offs, non-determinism — and the instinct to build the guardrails and evals that make them trustworthy
• A platform mindset: you build for leverage and clean interfaces, and you thrive in ambiguity in a fast-moving space

Desirable

• As a Staff AI Platform Engineer, you build the AI and data platform that powers Afresh's products: the knowledge and retrieval layer that makes grocery data reliably usable by LLMs, the agent systems built on top of it, and the evaluation and serving infrastructure underneath
• Your "customers" are Afresh's own engineers and AI products — your job is to give them a platform that turns raw grocery data into context a model can be trusted with, at production quality
• This is senior, 0-to-1 platform work
• You'll make foundational choices about how we represent grocery knowledge, ground our models, and measure whether any of it is actually working — in a fast-moving space with no playbook
• Build the knowledge & retrieval layer:
• Design and operate the knowledge graph and ontology that capture how grocery data relates
• Build the retrieval systems (vector, graph, and structured) that feed the right context to our models — so grounding is reliable, not lucky
• Build and serve the agent platform:
• Build LLM-powered agents (tool-use, multi-step reasoning, orchestration) and the serving infrastructure to run them reliably and cost-effectively
• Build the tools, abstractions, and interfaces other engineers depend on — a platform, not one-off features
• Own evaluation, quality, and the data foundation:
• Stand up eval sets, LLM-as-judge harnesses, tracing, and observability, plus the metrics (faithfulness, accuracy, hallucination rate, latency, cost) that tell us whether a change helped or hurt
• Build the pipelines, data products, and experimentation on Databricks/MLflow that take work from prototype to production — and partner with the engineers deploying our AI in the field to harden what works into reusable capabilities
• Our Tech Stack:
• Python, PySpark, dbt
• Databricks (Delta Lake, Unity Catalog, MLflow)
• Astronomer (Airflow) for orchestration
• LLMs and agents (Claude), retrieval over vector + graph stores, MCP tooling, eval and observability harnesses
• Claude, GitHub, Shortcut, Notion for development workflows