Staff Research Engineer, Agent Evals & Post-training

United States·San Francisco,London,Seattle+3 morefull-timelead

OtherResearch Engineer

2 views0 saves0 applied

Apply Now

Quick Summary

Key Responsibilities

Build eval suites for our agents, the MCP tool layer, and our internal company agent, grounded in real hardware tasks. Invent new benchmarks for agents working over physical engineering data,

Technical Tools

OtherResearch Engineer

Our mission is to accelerate how the world engineers new hardware. Nominal's connected test and operations platform powers the world's most advanced hardware programs and its most ambitious startups, from spacecraft, racecars, and autonomous vehicles to next-generation defense and energy programs. Our customers include Anduril, Shield AI, Hermeus, Albedo, Shinkei, and Pratt Miller Motorsports, as well as U.S. Navy and U.S. Air Force programs. Now we're expanding across the entire hardware lifecycle, building the foundation, AI-native applications, and agents that accelerate innovators' work and change what's possible to build.

We're backed by Sequoia, General Catalyst, Founders Fund, Lux Capital, and Lightspeed, and our team comes from SpaceX, Apple, Palantir, Anduril, Applied Intuition, and other leading companies.

We're the team behind Nominal's agents, AI-native applications, and MCP, and its forward-leaning AI bets. Our mission is to unlock the bottlenecks of the hardware lifecycle with AI. Our agents reason over physical reality, from high-rate telemetry and test campaigns to designs and simulations, where real test results are the ground truth their work is checked against. We believe opinionated AI, built for the real work of hardware programs, will change how the world engineers.

We're collaborative, iterative, and high-agency, and we're human-centered and customer-focused. We build with the newest AI tools every day, and because those tools keep changing, so do we: we stay curious and keep looking for the better way. Our team spans data science and ML, distributed systems, search, and knowledge systems, and we obsess over how agents can be genuinely useful to the engineers who rely on them.

As a Staff Research Engineer supporting agent evals & post-training, you'll define how Nominal measures its agents, for customers and for ourselves, and build the path from evals to post-trained models when hardware needs them. Evals come first; post-training follows when air-gapped deployment or cost makes it the right investment.

Responsibilities

~1 min read

→Build eval suites for our agents, the MCP tool layer, and our internal company agent, grounded in real hardware tasks.
→Invent new benchmarks for agents working over physical engineering data, where none exist today.
→Benchmark our agents against frontier agents using our MCP, and show where and why ours win.
→Build the eval infrastructure: datasets, model-based and human graders, and regression gates in CI.
→Turn evals into reward signals and training data, and lead post-training (fine-tuning, RL, distillation) when we need models we can run anywhere, including air-gapped environments.
→Set Nominal's strategy for measurement and model improvement, working with hardware experts on what "correct" means.

8+ years in ML engineering or research, including evals or post-training work you led in production.
Statistical rigor: you design evals that don't fool you, and you know when a difference is real.
Deep experience evaluating LLM or agent systems, including model-graded evals and their limits.
Hands-on post-training experience: fine-tuning, RL from feedback, or distillation on real tasks.
The judgment to know when to measure, when to train, and when a better prompt or tool is the answer.
A track record of setting technical direction across a team and raising the bar for the engineers around you.
You build with modern AI coding agents (Claude Code, Cursor, Codex) every day, and stay curious and open to better ways of working. The tools keep changing, and so do we.

Nice to Have

~1 min read

You've built evals or post-training at a frontier lab or an AI-native company.
You've published benchmarks or eval methods that others use.
You've trained or served open-weight models in restricted, on-prem, or air-gapped environments.
You've worked in test, reliability, or verification engineering for physical systems.