Lead Agentic RL Engineer for Inference-Time Optimisation

Recurvia; Greater London; On-site; GBP 90,000 - 150,000; Full time

Benefits offered by this job

Competitive salary

Equity participation

Job summary

Recurvia in Greater London is seeking an experienced Agentic RL Engineer to lead the agentic core of Metrana and contribute to a stealth project in inference-time optimisation.

You will own the Metrana agent end to end—designing the agent harness, building evaluation frameworks, running RL-based post-training, and developing autonomous research capabilities to investigate and act on training workloads with minimal supervision.

Qualifications

• MSc, PhD or equivalent in CS/ML.
• Solid RL theory background; apply to post-training of LLMs.
• Hands-on with GRPO, RLVR, RLHF.
• Experience designing agent harnesses and evaluation loops.
• Production-grade Python and PyTorch experience.
• Ability to drive research and engineering projects independently.

Responsibilities

• Design and build the Metrana agent harness with MCP integration.
• Develop rigorous evaluation frameworks for agents.
• Run RL-based post-training using GRPO/RLVR and related techniques.
• Build autonomous research agents for experimentation and analysis.
• Contribute to a stealth project on inference-time optimisation.
• Develop tool-use, code, and computer-use agent capabilities with long-horizon planning.
• Collaborate to ship agentic systems into production.

Skills

MSc/PhD or equivalent

RL theory

GRPO/RLVR/RLHF

Agent harness design

Evaluation framework design

Python

PyTorch

Independent project leadership

Education

MSc, PhD or equivalent

Tools

MCP protocol

Python

PyTorch

Job description

Recurvia in Greater London is seeking an experienced Agentic RL Engineer to lead the agentic core of Metrana and contribute to a stealth project in inference-time optimisation.

You will own the Metrana agent end to end—designing the agent harness, building evaluation frameworks, running RL-based post-training, and developing autonomous research capabilities to investigate and act on training workloads with minimal supervision.