This role is with one of Dex’s trusted partner companies. We work closely with their teams to truly understand their culture, goals, and what they’re looking for, so we can match you with the right opportunity and give you context about the role before you commit to a process.
Dex is an AI recruiter agent that helps you run your job search. Tell Dex your stack, seniority, and what you want to build. We will manage your applications and surface other opportunities that are a fit.
The role
While frontier AI labs race on parameter scale, the core bottleneck to autonomous agents is verifiable environments where models can act, fail, and learn from real software feedback. Backed by top-tier technical investors and led by a repeat founder with a prior exit, this fast-growing lab builds interactive, resettable digital sandboxes purpose-built for agent training. Within their first year, they have scaled to over 50 staff across London and Paris and are already embedded directly with leading frontier research labs.
As a Member of Technical Staff, you will take ownership of the full post-training and harness loop, sitting squarely between novel research and hard systems engineering. This is neither a theoretical paper-writing post nor an application-layer seat wrapping commercial APIs. You will formulate your own technical bets, curate large-scale training distributions, fine-tune models using SFT and RL, and build the sandboxed execution harnesses required to test agents against complex digital tasks. Scope spans Senior to Principal levels based on technical depth and systems judgement.
The work
- Design and execute end-to-end post-training pipelines, applying supervised fine-tuning (SFT) and reinforcement learning (RL) to train agents on complex multi-step digital tasks.
- Architect the sandboxed execution environments, agent harnesses, and deterministic feedback loops that allow models to interact safely with real-world software.
- Build and scale the underlying training-data infrastructure, converting messy digital workflows and synthetic data into curated datasets that move benchmark performance.
- Establish rigorous evaluation frameworks to benchmark agent capabilities, diagnose failure modes, and guarantee reproducible model iterations.
- Provide technical leadership across ambiguous research directions, translating open-ended capability goals into production-grade systems alongside curriculum and software engineers.
What You Bring
- Demonstrated track record directly training, fine-tuning, or post-training models (via SFT, RL, or RLHF) where you can speak to data regimens, loss dynamics, and performance deltas—not just prompting external model APIs.
- Deep foundational knowledge of machine learning fundamentals, including model optimization, reinforcement learning architectures, and rigorous evaluation methodologies.
- Production-grade Python engineering ability with the discipline to build modular, tested, observable systems rather than one-off research notebooks.
- Demonstrated history of taking ambiguous, first-principles problems from raw concept to working system over multi-month horizons without relying on established playbooks.
- Based in or willing to relocate to London or Paris.
#J-18808-Ljbffr…