Research / Data Engineer

Emulated builds high-quality, long-horizon reinforcement-learning environments and post-training data for frontier labs. Customers include the major US frontier labs and top international labs. Most data vendors treat post-training data as a staffing problem. Emulated treats it as a research problem: scale faster than headcount by building a model that can produce post-training data on its own.

It is cash-flow positive. It raised a sub-$1M pre-seed, then took a seed offer a few months later to fund model training. The team is about 9 to 12 people: a published neuroscience research lead, founders with AWS software backgrounds in distributed databases and network infrastructure, people from Google, DeepMind, and NVIDIA, and a Stanford shock-physics PhD. They recently released Auto Research, a benchmark of how well models iterate in a machine-learning research loop. Seed stage, founded in 2025, in San Francisco.

Emulated is hiring a research and data engineer because inbound demand for post-training data is bigger than the team can fulfill. The shape that works is a domain or research expert with a real ML or engineering spike. You do the research and you produce the data. For this hire those two jobs are one seat.

You post-train models end to end, from provisioning infrastructure through curating high-quality data. You build RL environments and long-horizon tasks across domains, including research loops, science, chips, and shock physics. The point is to scale data generation faster than you add people, through self-improving processes. Data taste matters: judgment about which problems in a domain are real, and which tasks a lab will actually buy.

Bar: about 3 to 7 years. A genuine spike in post-training, RL, evals, interpretability, or similar, plus strong software engineering. You can post-train a model and curate the data it learns from. Fast code comprehension. The technical round is a self-contained problem, not LeetCode. The ML version uses PyTorch and debugging ML pipelines. Publications, side projects, open source, or a blog are the signal they look for. Domain depth in neuroscience, physics, chemistry, or systems fits. Palantir, Anthropic, or OpenAI is a soft signal. Competitor data companies are interesting but less defined. A data-services mindset, or only one of research and engineering, is a miss.

In office in San Francisco. Visa sponsorship is available. Posted work intensity is 9-9-6. The process is an intro with a founder, a technical round, then a work trial.