Research Engineer, Reinforcement Learning

Axiōma Search; Greater London; In loco; GBP 90.000 - 130.000; Tempo pieno

Descrizione del lavoro

Seed-stage AI for Science Lab in London (on-site) is seeking an engineer to shape training setups for RL-driven agents in biology and drug discovery. You will design experiments, build environments, and work with scientists to turn open biological questions into testable tasks.

You will contribute to end-to-end experimentation, debugging, and reproducible results, while helping set tooling and research standards in a small, ambitious team.

Competenze

• Hands‑on experience training models or agents with RL
• A first‑principles grasp of reward design, credit assignment, exploration and training stability
• Strong engineering - you write reliable, fast code and can debug from learning dynamics down to GPU performance
• Sound experimental judgement: you know why a run failed and what to try next
• Comfort owning ambiguous problems in a team of a handful of people
• Willingness to work five days a week in the London office

Mansioni

• Train agents with RL on multi-step problems across biology, drug discovery and clinical research
• Build environments where agents query biological data, call scientific tools and act on the output
• Design reward functions and evals that separate real progress from reward hacking
• Build rollout, distributed training and eval infrastructure so a few people can run large experiments
• Own experiments end-to-end, from design through debugging to reproducible results
• Work with in-house scientists to turn open biological questions into tasks with checkable outcomes

Descrizione del lavoro

Seed-stage AI for Science Lab | London (on-site)

Overview

RL has pushed coding and maths forward quickly. Biology is harder because evidence is incomplete, feedback is slow, and a model can look like it's improving while quietly gaming the reward. This role exists to build training setups where that can't happen.

The lab is a small, seed-funded founding team in London. They train frontier models as agents on real scientific work: forming hypotheses, planning experiments, reading results.

You'd be one of the first technical hires. There's no inherited stack, so you'd help set the tooling and research standards.

What you'll do

• Train agents with RL on multi-step problems across biology, drug discovery and clinical research
• Build environments where agents query biological data, call scientific tools and act on the output
• Design reward functions and evals that separate real progress from reward hacking
• Build rollout, distributed training and eval infrastructure so a few people can run large experiments
• Own experiments end-to‑end, from design through debugging to reproducible results
• Work with in-house scientists to turn open biological questions into tasks with checkable outcomes

What you'll need

• Hands‑on experience training models or agents with RL
• A first‑principles grasp of reward design, credit assignment, exploration and training stability
• Strong engineering - you write reliable, fast code and can debug from learning dynamics down to GPU performance
• Sound experimental judgement: you know why a run failed and what to try next
• Comfort owning ambiguous problems in a team of a handful of people
• Willingness to work five days a week in the London office

Optional Bonus

• Agentic systems work: tool use, multi-step reasoning or long‑horizon tasks
• Biology or healthcare background - useful, not expected
• PhD or publications in RL, post‑training or scientific ML

Seniority is open, from strong early‑career engineers to senior researchers.

Shortlisted candidates will be contacted within 48 hours.