Research Engineer, Training and Environment Infrastructure

Ersilia; San Francisco (CA); On-site; USD 150,000 - 350,000; Full time

Benefits offered by this job

Meaningful equity grants

Health, dental, and vision coverage

Job summary

Ersilia in San Francisco is building the data and infrastructure behind frontier model training and deployed AI agents. You will own evaluation systems, ranking training tasks by value and catching drift as data and tasks evolve.

This role is hands-on and in person, six days a week, requiring strong systems engineering, RL experience, and the ability to ship production-grade evaluators tied to training pipelines.

Qualifications

• Experience building environments, evaluations, or RL pipelines in production.
• Hands-on RL research or data science experience.
• Strong systems engineering skills and the ability to ship end-to-end evaluators tied to training pipelines.

Responsibilities

• Rank training value before spending compute and build a system to rank tasks by expected training value.
• Design scoring that runs cheaply at scale, tuned to customer-specific outcomes.
• Detect and adapt to distribution drift by updating evaluations as customer requests evolve.

Skills

RL pipelines

Systems engineering

Reward design

Ambiguous evaluation problems

Job description

About the company

We work on the data and infrastructure behind frontier model training and deployed AI agents. We partner with AI labs and large organizations on post-training research, infrastructure, and deployments. We are an early-stage, revenue-generating team hiring for in-person roles in San Francisco.

About the role

You sit at the center of the post-training loop. Before compute is spent on a dataset, you decide whether it is worth it. After a model change ships, you determine whether it helped.

The work spans reward design, automatic scoring at scale, and staying ahead of drift as customer traffic and task shapes evolve. This is not a fixed benchmark you maintain once and forget.

What you'll do

Rank training value before spending compute

Work out which tasks in a dataset are worth training on, and build a system that ranks every task by expected training value.

Build scoring that runs cheaply at scale

Design automatic scoring cheap enough to run constantly, tuned to each customer's definition of a good outcome rather than generic correctness.

Catch drift before it becomes a problem

Notice when real customer requests have moved far enough from the existing test set that it no longer describes the job, then rebuild the evaluations to match.

What we're looking for

• You have built environments, evaluations, or RL pipelines that ran in production.
• You have built RL data or done hands-on RL research, rather than working adjacent to it.
• You are a strong systems engineer who can pick up any stack.
• You have substantive opinions on reward design, including what makes a checker trustworthy and how models learn to game them.
• You are comfortable owning ambiguous evaluation problems end to end, from building the first version and inspecting the data to iterating until the system is useful.
• You can work six days a week, in person, in San Francisco.

Nice to have

• Experience with LLM-as-judge systems, automated graders, or rubric design for model outputs.
• A background in applied ML research, research engineering, or data science at a lab or fast-moving startup.
• Experience building evaluation infrastructure tied directly to a training pipeline, not just offline benchmarking.

Compensation and benefits

$150,000 to $350,000 USD base, depending on experience and seniority

• Meaningful equity grants for early team members
• Health, dental, and vision coverage