Founding applied AI engineer in causal inference

Referment is working with an early-stage, venture-backed AI company whose platform turns enterprise data and live market signals into governed, ranked recommendations for strategic decisions — the kind a board signs off on. Every recommendation carries its evidence and its limits, each decision is logged so the business can later prove what worked, and the platform runs entirely inside the client's own cloud environment. They're now setting up an applied AI research lab, and this role is one of its founding seats.

• You'll own the journey from a client's decision to a proven cause.
• Building the causal layer of the platform: map what drives each client's numbers, estimate effects from natural experiments, and refute every estimate before it's used
• Returning effects as honest ranges — how much, and with what lag — so nothing prescriptive ships until it's calibrated
• Designing experiments with clients, such as a price change on a single segment, and turning forecasts and causal estimates into "if we do X, here's the range"
• Owning the attribution method that credits each decision with the result that followed
• Designing the evaluation that decides what ships: benchmarks, scoring rubrics and blind holdouts, plus the statistics that make them trustworthy
• In your first 90 days you'll learn the platform on real client questions, ship a first causal estimate with honest ranges on a live client problem, and move the evaluation from a single quality score to rubric-based checks with a blind holdout.

• Around five or more years building AI or ML systems that run in production
• Causal inference in practice — this is non-negotiable. You've estimated the effect of a real intervention on business data (difference-in-differences, synthetic control, uplift, marketing-mix models or experiments) and you know when each method breaks on short histories
• Strong statistics: uncertainty, calibration and experiment design
• Evaluation design: you've built benchmarks that decided what shipped, and you know how to check that the evaluation itself can be trusted
• Very strong Python (async, typed, tested) and solid SQL, with systems you've deployed and maintained on AWS or Azure
• Real LLM depth: at least one RAG, tool-using or text-to-SQL system taken properly beyond a demo. You know LangChain, LangGraph and LlamaIndex, and when you're better off writing it yourself
• AI-native working habits — tools like Claude Code in daily use — and you own every line you merge
• A degree from a top-tier institution, with AI-native training and learning
• You write your design down before you write code, and you can explain cause and effect to a CFO in plain English
• Будет плюсом: time series forecasting; Bayesian methods; fine-tuning; ML served and monitored in production; Databricks; published research or open-source work; experience leading a small team

• The role is remote and UK-based, full-time, with founding-team equity.