Role: Member of Technical Staff - Evaluations
Team: Post-Training
Location: San Francisco
Location Type: Hybrid
Employment Type: Full-time
ABOUT LIQUID AI
Spun out of MIT CSAIL, we build general-purpose AI systems that run efficiently across deployment targets, from data center accelerators to on-device hardware, ensuring low latency, minimal memory usage, privacy, and reliability. We partner with enterprises across consumer electronics, automotive, life sciences, and financial services. We are scaling rapidly and need exceptional people to help us get there.
THE OPPORTUNITY
Public benchmarks saturate and don't generalize to real-world tasks, so post-training risks overfitting to external scores. Every post-training decision at Liquid AI needs internal evals we trust: evals that track real-world behavior and are cheap and reproducible to run. This role sits in the Post-Training team and exists to make sure those decisions rest on exactly that kind of evidence.
You will work on the internal eval suite and infrastructure end to end working closely with the post-training researchers whose decisions depend on it and with other teams (audio and beyond) who need evals for their own models. Evaluating small edge models is an open problem, since nearly every public eval targets frontier models, and the evals you build here change the quality and capabilities of models we ship within weeks, not half a year.
WHAT WE'RE LOOKING FOR
We need someone who:
- Proactive execution: You pick up use cases and come back with working benchmarks. You parallelize work, make progress visible, and finish without being chased.
- White-box about evals: You volunteer where your own evals or scores were wrong and what you changed, and you present numbers with their caveats attached.
THE WORK
- Work on the internal eval suite and infrastructure: keep it reliable, optimized, and reproducible at scale.
- Design held-out evals for priority capabilities such as agentic workflows, computer use, and cybersecurity, with documented construction and contamination checks.
- Build evals targeted at small edge models, where most public evaluations built for frontier models don't apply.
- Read generated outputs to debug unexpected scores before anyone acts on them.
- Turn customer and open-source feedback into environments usable for data, training, and evaluation, and partner with other teams to design evals for their specific needs.
- Release open-source evals, write papers about them, and publish failure-mode reviews of popular evaluations.
EXPERIENCE & CAPABILITIES
What you'll need:
- Eval design is where we need you to be unusually strong: you can take a high-level description of a capability and turn it into an evaluation that is accurate, general-purpose, separates good models from bad ones, and is hard to game.
- Evals you have built before that we can recognize or check, and that hold up on coverage, reproducibility, contamination handling, and the signal they were meant to give. At least one built from scratch: a benchmark, an eval paper, or real contributions to lm-eval-harness, lighteval, OLMES, HELM, or Inspect.
- Fluency with the fundamentals: contamination, saturation, judge bias, and variance.
- Hands-on experience running and managing evals in a team, whether on an eval team, a training team, or in an engineering/DevOps role.
- The habit of debugging unexpected scores by reading individual generated answers before believing the number.
- Software engineering strong enough to own eval infrastructure and debug the inference and infra failures that surface when running evals at scale, without hand-holding.
Helpful, not required:
- Model training experience.
- Publications or research background (industry backgrounds are equally welcome).
- Deep systems expertise.
- Familiarity with what small edge models do and don't do well.
- An eval-specific title is not needed: training-team or DevOps profiles who ran a lot of evals are a good fit.
WHAT SUCCESS LOOKS…