Research Engineer, Reinforcement Learning.
- Location
- London
- Working pattern
- Hybrid from January 2027
- Compensation
- £120K + bonuses + equity
Most RL environments are simulations. Ours are not.
For two years we have recorded real commercial decisions as they happened in production, from pre-market to won deal. Each trajectory carries the state at every step and labels across 20 or more sequential decision points. Labels are emitted by our production applications and verified three ways: by real outcomes, by the accountable user, and by our own operators.
Our models are being trained for initial deployment post logging epoch completing in January 2027.
FBED Labs is completing its seed round. In 2027 we focus on two things: training regimes for our own model ladder, and task suites and RL environments for training partnerships with AI labs. We are hiring the engineer who will lead that research day to day.
What you’ll do
- ·Build long-horizon environments and task suites from real sequential decision data
- ·Design state representations that stay sufficient across long, partially observed trajectories
- ·Train, post-train and evaluate our own models, from parsing models to Shannon Zero
- ·Run rigorous evaluations against human raters and frontier models
- ·Own the live policy and the labelling pipeline, so data collection meets the coverage that training and off-policy evaluation require
- ·Run QA on environments delivered to lab partners and be their technical counterpart
- ·Explore open problems in model memory: what to keep, what to consolidate and what to forget
- ·Lead a second research engineer, working with our CTO and CEO and alongside an ML platform engineer
What you’ll bring
- ·Hands-on experience in reinforcement learning or sequential decision modelling
- ·Experience learning from logged data: off-policy evaluation, importance weighting and its diagnostics, such as effective sample size
- ·A track record on long-horizon credit assignment
- ·Working knowledge of causal inference and causal graphical models
- ·Strong experience training and evaluating models (PyTorch or similar)
- ·Experience designing environments, benchmarks or evaluations
- ·Comfort with messy real-world data and the judgment to make it usable
Nice to have
- ·Post-training, reward modelling or offline RL
- ·Experience building data or environments for model developers
- ·Interest in memory, continual learning or long-context systems
Why FBED Labs
Our commercial operation funds the lab. You get real production data, real customers and a direct line from research to deployment, without waiting on a grant cycle or a simulator.
Apply below with papers, a repo, or a short note on a credit assignment problem you have worked on.