Machine Learning Engineer – Lumen Enterprise Models (SWE-focused LLMs)

Overview

Cosine builds autonomous software-engineering systems for organisations where reliability, security and control matter most. Its Lumen Enterprise models are designed to work across real codebases, understanding unfamiliar architectures, changing multiple files, using tools, running tests and producing maintainable software. The company is hiring a full-time Machine Learning Engineer in London to help train and improve these models. The role reports to the CEO, starts as soon as possible, and is in the office full time by default.

The engineer works across the whole training stack, including pre-training, supervised fine-tuning and reinforcement learning, and should be equally happy profiling a training loop for bottlenecks and memory leaks or implementing new RL algorithms. They work with research, infrastructure and product on questions such as what the model should learn next, which data and environments teach best, how to reward correct and maintainable software, and how to tell when a model is genuinely better for engineers.

Tasks include running pre-training, fine-tuning and reinforcement-learning experiments, optimising training at scale on frontier models, fixing performance and communication bottlenecks in distributed training, training on long-context, multi-file and repository-level tasks, building RL and RLVR loops using signals from execution, tests, linters and static analysis, and investigating failure modes such as poor planning, brittle fixes and reward hacking.

Requirements

Cosine seeks strong experience training deep-learning models, deep practical experience with PyTorch or a comparable framework, experience with distributed and multi-GPU training such as FSDP or DeepSpeed, strong software-engineering fundamentals and clear communication. Meaningful work on at least one of pre-training, SFT, preference optimisation, RLHF or RLVR is important. Mixture-of-experts, CUDA, vLLM or SGLang, SLURM or Kubernetes, code-model benchmarks such as SWE-bench and open-source contributions are bonuses.

What you get

Compensation is a market-benchmarked salary with equity options available for discussion, 30 days holiday plus bank holidays, a pension, quality equipment, a dog-friendly office, weekly team breakfast and lunch, and monthly socials. The team works in the office five days a week.

Key dates

The application deadline is 8 November 2026 at 23:59 UK time.

Who can apply

Visa sponsorship is available for suitable candidates. Cosine encourages applications even where not every qualification is met.