AI Research Engineer, Post-Training

Lovable; Greater London; On-site; GBP 120,000 - 180,000; Full time

Job summary

Lovable in London seeks an engineer who has hands-on post-training experience at scale to translate research into production training recipes and run models in production for code generation and agent workloads.

You will own the full post-training pipeline, applying RL, preference optimization, and supervised fine-tuning to improve model behavior and reliability. You’ll work on production systems, GPU orchestration, and data pipelines to ship fast.

Qualifications

• Experience running post-training jobs on large language models.
• Solid production code ability for reliable systems.
• Experience with distributed training and GPU clusters.

Responsibilities

• Own the full lifecycle of Lovable's post-training pipeline from data curation to deployment.
• Apply reinforcement learning, preference optimization, and supervised fine-tuning to improve code generation and agent performance.
• Build evaluation and experimentation infra to measure real-world impact of changes.
• Develop and operate production systems that run training jobs at scale, including GPU orchestration and data pipelines.
• Collaborate with agent, product, and infrastructure engineers to translate model gains into user-visible improvements.
• Investigate failures end-to-end across training, data, and serving.
• Read papers, run experiments, and move quickly to production.

Skills

Post-training at scale

Production-grade code

Distributed training

GPU clusters

Evaluation systems

Model debugging

Tools

PyTorch

JAX

Job description

TL;DR:
Lovable lets over 2 million people build software using plain language, and the models behind it need to be exceptional. We're hiring an engineer who has gotten their hands dirty with post-training at scale and wants to do it again for one of the fastest-growing AI products in the world.

You’ll own our full post-training pipeline: translating the latest research into production training recipes, adapting them for code generation and agent workloads, and putting improved models in front of users fast. The goal is to get promising research into production within days or weeks, not months. This isn't an academic research position - you'll spend as much time in production infrastructure as in training configs, and your success is measured by what ships.

Why Lovable?

Lovable is the software creation platform that gives people the power to act on the problems closest to them. For decades, turning an idea into software required so much capital, technical fluency, and time that many ideas never came to life. Lovable is the counterargument: a platform for all people with ideas, ambition, and problems worth solving. From solopreneurs to small business owners to teams at companies like Adidas and Zendesk, people have built over 60 million projects on Lovable since its launch in November 2024. And we’re just getting started.

We’re building a generational company from Stockholm, with growing teams in London, Boston, New York, and San Francisco. Our team is small, talent-dense, and moving quickly, with a culture rooted in extreme ownership, high velocity, and low-ego collaboration. We look for people who care deeply, ship fast, and are eager to make a dent in the world.

Lovable is one of TIME's 100 Most Influential Companies and has been recognized on the Forbes AI 50 and CNBC Disruptor 50, reflecting our momentum as one of Europe’s fastest-growing AI companies and one of the most ambitious places to build in this next era of software.

What we’re looking for

• You’ve personally run post-training jobs on large language models - RFT/RLVR, preference optimization, or similar. Not just called APIs or written prompts, but actually trained and iterated on models
• You can write solid production code. The systems you build need to run reliably, not just produce interesting research artifacts
• You're fluent in at least one major ML framework (PyTorch, JAX) and comfortable working with distributed training setups and GPU clusters
• You understand the math behind preference optimization, reward modeling, and alignment techniques - and can reason about when each approach fits
• You’ve built or significantly contributed to evaluation systems that capture real-world quality, not just benchmark scores
• You can trace a model quality regression from user-facing symptoms back through serving, inference, and training - and you enjoy doing it
• You want to ship. Research taste matters, but at Lovable the question is always "how fast can we get this to users?"

Preferred:

• You’ve worked on code generation or agentic use cases specifically
• You’ve put post-trained models into the hands of real users and seen how they hold up at scale
• You’ve owned the full loop: curating data, running training, evaluating results, deploying, and monitoring in production
• You have a habit of reading a paper on Monday and having a prototype running by Friday
• You’ve experimented with speculative decoding or similar techniques to improve model efficiency
• You have strong views on evaluation methodology and have built evals that actually predict user satisfaction
• You’ve published or contributed meaningfully to the open-source ML ecosystem

What you’ll do

• Own the full lifecycle of Lovable's post-training pipeline - from data curation and training runs through evaluation and deployment
• Apply and adapt reinforcement learning, preference optimization, and supervised fine-tuning methods to make our models better at generating code, reasoning about user intent, and acting as reliable agents
• Build the evaluation and experimentation infrastructure that tells us whether a model change actually helps users - covering helpfulness, safety, latency, and reliability
• Develop and operate the production systems that run training jobs at scale, including GPU orchestration and data pipelines
• Work across team boundaries with our agent, product, and infrastructure engineers to turn model gains into product improvements users can feel
• Investigate and resolve failures end-to-end - whether the root cause is in a training recipe, a data issue, or a serving regression
• Read papers, run experiments, and move fast: the goal is to get promising research into production within days or weeks, not months

We treat all candidates equally