Member of Technical Staff, Applied RL

Stealth - AI x Bio; Greater London; In loco; GBP 120.000 - 180.000; Tempo pieno

Vantaggi offerti da questo lavoro

30 days annual leave

Private health cover (UK)

Pension / UK

Descrizione del lavoro

Stealth - AI x Bio is hiring a role focused on teaching our agents to learn through reinforcement learning. You will design harnesses and long-horizon tasks, train LLMs with RL, and build robust evaluation methods, often bridging biology and AI teams in London or San Francisco.

You will work with scientists and engineers to ensure careful reward design, detect reward hacking, and advance agent capabilities using Python and PyTorch in a collaborative, open development environment.

Competenze

• Hands-on experience training LLMs with reinforcement learning, using methods such as PPO or GRPO, including reward design and detecting reward hacking.
• Experience with RL training frameworks such as veRL, Miles or TRL, RL environment frameworks such as Harbor or Prime Intellect's verifiers, and inference engines such as vLLM or SGLang.
• Building agent harnesses and long-horizon tasks with tool use, such as search, code execution and external APIs.
• Designing learning signals for long trajectories, including verifiers, LLM judges, process reward models and credit assignment across steps.
• Rigorous comparisons, with matched compute, repeated runs, noise estimates, ablations and held-out problems.
• Clear writing and communication about experiment design, results and what they do and do not show.
• Strong Python and PyTorch, and daily use of coding agents.

Mansioni

• Own how our agents learn, including harnesses and RL environments.
• Design and implement tasks, rewards, and evaluations for long-horizon agent learning.
• Collaborate with Bio ML researchers, evaluation and infrastructure engineers, software engineers and scientists.
• Assess when improvements come from models vs. harness/tools/tests.

Conoscenze

LLM training with RL

PPO/GRPO reward design

RL frameworks (veRL, Miles, TRL)

Python

PyTorch

Agent harnesses and long-horizon tasks

Tool use: search, code execution, APIs

Experiment design and evaluation

Communication of results

Strumenti

veRL

Miles

TRL

Harbor

Verifiers

vLLM

SGLang

Descrizione del lavoro

The opportunity

We're a stealth startup in AI and bio, building infrastructure for data generation. The team is small and elite, the problems are hard, and the foundations are being laid right now. You'd help write them from the first line, with real ownership and a direct line to the founders.

About us

AI for biology has a data problem. The data that matters most doesn't exist yet, so we're building the infrastructure to produce it, with quality and traceability built in from day one. We're in stealth and heads-down on execution. We'll share more about what we're building once you've spoken with the team.

The role

You will own how our agents learn, which covers building the harnesses and RL environments they work in and post-training the models that run through them. Our agents work on long-horizon scientific problems, reading papers, querying databases, running code and calling biological models to decide what to test next, and they also need to notice when an experiment has gone wrong. You will also build environments where agents carry out experimental work, not only plan it.

Rewards in science are noisy, expensive and slow to arrive, so careful reward design matters more than the scale of a run. When an agent improves, you will work out whether that came from the model itself or from a change in the harness, the tools or the task.

You will report into the intelligence team and work closely with the Bio ML researchers, the evaluation and infrastructure engineers, our software engineers and our scientists.

Who you are

You have trained or post-trained language models and agents with reinforcement learning, and built the tasks, rewards, harnesses and evaluations around them. You read the traces before trusting an average, and you start from a baseline you trust. A background in biology is not required, as you will be working closely with our scientists.

MUST HAVE

• Hands-on experience training LLMs with reinforcement learning, using methods such as PPO or GRPO, including reward design and detecting reward hacking.
• Hands-on experience with RL training frameworks such as veRL, Miles or TRL, RL environment frameworks such as Harbor or Prime Intellect's verifiers, and inference engines such as vLLM or SGLang.
• Building agent harnesses and long-horizon tasks with tool use, such as search, code execution and external APIs.
• Designing learning signals for long trajectories, including verifiers, LLM judges, process reward models and credit assignment across steps.
• Rigorous comparisons, with matched compute, repeated runs, noise estimates, ablations and held-out problems.
• Clear writing and communication about experiment design, results and what they do and do not show.
• Strong Python and PyTorch, and daily use of coding agents.

NICE TO HAVE

• Experience applying RL or agents to biological problems, for example with frameworks such as ProtRL.
• Experience with multi-agent setups, such as proposer and critic agents or debate.
• Experience learning from noisy or delayed rewards, and calibrating model confidence.
• Publications, open-source work or public benchmarks in RL, agents or post-training.

How we work

The role is based in London or San Francisco. We work in the open, write decisions down, and expect each person to own their area and make the calls inside it.

Everyone gets 30 days of annual leave plus local public holidays. In the UK that comes with a pension with a 10% employer contribution and Bupa private health cover; in the US, a 401(k) with matching and health cover. We add to these as the team grows.

Our process

Four stages: a screening call, a behavioural interview, an in-person technical session with the team on a real problem from our work, and references.

If you are not sure you fit, apply anyway.

We are an equal opportunity employer. We make hiring decisions on merit, scope-fit, and the strength of the working relationship we expect to build with each hire. Applications welcome from candidates of any background.