Inference research

Founding Post-Training Researcher

Be our first dedicated post-training hire. You’ll own the loop that turns customers’ production traces into smaller, cheaper open models built for their workloads.

Location
Cambridge
Office
Mon, Wed, Fri
Compensation
Top-of-market compensation + significant equity

About Gradiated

  1. Gradiated is a research lab focused on inference. We co-design software and hardware for efficiently serving open-weight models.
  2. Our goal is to maximize intelligence per joule. We work across model behavior, software, systems, and silicon to find better ways to run models.
  3. We start by thinking about the model as a mathematical function, and map that out onto hardware and then build software on top of that which is custom to each model. This maximises performance and removes wasted abstractions.
  4. We are early. You will help set the technical direction and the way the company works.

Why this role exists

  1. Our automated RL product, Gradiated Training, is in private beta. It takes a team’s production traces, builds evals around their tasks, and post-trains an open model for that workload. In one early case, SFT followed by GRPO took GLM-5.3 Flash from 54% to 76% on a support agent task, at about 1/25th the cost of a frontier model.
  2. The loop is starting to work, but it’s early. You’ll be our first dedicated post-training hire. You’ll own and improve the loop that takes production trace data and turns that into cost effective LoRA adapters.

What you will do

  1. The pipeline, end to end: take raw production traces all the way to a model we’re confident putting into production
  2. Recipes that generalize: Test SFT, GRPO and new training methods across real customer environments. This will be hands on, working with customer data to see what works, and then trying to generalize it
  3. Building an RL environment pipeline: We are using customers trace data and github repositories to reconstruct RL environments. This has a lot of variation in practice, so working directly with customer data to trial this on many different environments is important.
  4. Honest evals: judge models, rubrics and baselines that tell a customer what got better and what got worse
  5. The team: help hire and shape post-training as it grows, and share what we learn publicly

What we are looking for

  1. Several years post-training large language models that shipped to real users, for example at DeepMind or a frontier lab. You owned the outcome, not just the experiments
  2. Hands-on experience running RL for LLMs (RLHF, PPO, GRPO, DPO or similar), and you have experience of how things break
  3. Experience building good evals
  4. You've built infrastructure for post-training, and run training jobs at scale in production
  5. You have experience leading a team focused on post-training
  6. Low ego. You write clearly, you ship, and you’d rather fix the thing than present about fixing it
  7. Based in London or Cambridge. We’re in person in Cambridge on Monday, Wednesday and Friday

Problems you might work on

  1. Reward design and judge reliability when every customer’s task is different
  2. Turning noisy production traces into useful training data, and automating that process across multiple customers
  3. RL on long-running, multi-step agent tasks
  4. Avoiding regressions on things the customer didn’t measure
  5. Making training runs cheap and fast enough to run continuously as traffic changes

This role isn’t for you if

  1. You want a well-defined role with little ambiguity, we are an early stage startup so you should be comfortable handling ambiguity in the work you're doing
  2. You are looking for a role to develop as a manager more than an IC, despite leading the post-training team, your will mostly be doing IC work
  3. You'd like to publish lots of papers - this role will involve writing publicly and sharing the research work you do, but the goal won't be publishing papers
  4. You've mostly used off the shelf tools for fine-tuning models, or one off training scripts, rather than post-training models in production
  5. Your career has been almost entirely at big tech, without time in the scrappy, little-exists-yet setting of a seed-stage startup
  6. It’s the other way round: you’ve only worked at small startups and haven’t yet operated systems at real scale
  7. A short stint is what you’re after. We want someone to build this with us for years

How we work

  1. We hire great people and get out of their way. You’ll be the one driving this, and you won't be micro-managed by someone
  2. We take ownership, which means making systems reliable so we’re not fighting fires in production
  3. We experiment a lot and take on hard problems, even when they’re likely to fail

Pay and equity

  1. We offer top-of-market compensation, with significant equity. If you are interested, but you think pay will be a blocker, please apply anyway
  2. We pay at the top of the market because talent density is what allows us to build the most efficient inference. We want a small team of exceptional people who have fun solving hard problems together.

Interview process

  1. A 60-minute call about your goals and the role.
  2. We get together on Zoom and talk through the problems we will work on.
  3. A paid day of working, so we can get to know you and you can meet the team.
  4. References and an offer decision.

Apply

Let's talk.

If this sounds interesting to you, we'd like to hear from you. Fill in the form below and we'll get back to you within a few days.

Loading the application form