Machine Learning Engineer, Model Experimentation

About the role

This role exists to test modeling ideas quickly and tell the team, with evidence, which ones deserve more time. The ML team is small and senior; this person is its fastest experimental loop.

A typical cycle looks like this: take an idea (a new architecture, a different representation, a change to the training objective), set up a baseline, run it, and come back with a clear call to continue, change course, or stop. Most ideas will not work, and that is expected. Success in this role is measured by how quickly and clearly you learn which ideas hold up, not by how many of them do.

You will work day to day with senior members of the ML team, who will help set priorities and review your calls. The data you work with comes from real lab measurements, so results can be checked against what actually happened at the bench.

What you'll do

• Design and run ML experiments end to end: baseline, model changes, training runs and evaluation.
• Make a clear call on each experiment and write it up in a page or less, including the ones that failed.
• Build evaluations you would trust even when they go against the idea being tested.
• Bring ideas from the wider field, such as world models, latent-prediction objectives or diffusion, and test them against Aether's own data.
• Work within the team's ML infrastructure, and improve it where it slows experiments down.

What we're looking for

• Hands-on experience training and evaluating models yourself, whether in research, a job, competitions or serious personal projects. The number of experiments you have run matters more than years of experience.
• Strong experimental habits: baselines first, 1 change at a time, ablations, and honest evaluation.
• Speed. You are comfortable writing code that runs today and cleaning it up only once an idea earns it.
• A track record of stopping things. You can describe an experiment you abandoned, what told you to stop, and how quickly you made the call.
• Clear written communication of results, especially negative ones.
• Fluency in Python and a modern deep learning framework such as PyTorch or JAX.

Nice to have, and what we'll teach

• Experience with world models or latent-prediction methods (JEPA, DreamerV3 or similar)
• Experience with diffusion models, in any domain
• Competition results, open-source contributions, or a project you built and maintained yourself

Note: No chemistry or biology background is required. We will teach the molecular and protein side, along with Aether's data and tooling. We would rather hire a strong experimenter and teach the domain than the reverse.