Research Scientist, Post-Training

San Francisco, CA · On-site · Full-time

Compensation: $150K–$250K base + profit sharing (total cash ~$250K–$450K) + equity

About the Company

A well-funded early-stage (post–Series A) company building high-signal training data and evaluation infrastructure for frontier AI labs, with a founding team drawn from top quant firms and leading AI labs. They partner with leading labs to design datasets and run rigorous evaluations that go beyond static benchmarks. Small team where individual contributors have direct impact on how the next generation of models learns and improves.

Founded 2025 · 11–50 people · Industry: AI / ML — training data & evaluation infrastructure

The Role

Prove that the data works — design and run training experiments that isolate the impact of the datasets on model behavior (SFT and RL post-training), and turn the results into defensible evidence for partner labs. Experimental, high-leverage IC work at the edge of model development.

What you'll be doing

• Run controlled SFT and RL experiments to measure the impact of the datasets on model performance.
• Quantify lift across capabilities — reasoning, tool use, long-horizon tasks, and domain-specific workflows.
• Share findings directly with partner labs to deepen relationships and drive sales.
• Collaborate with internal SPLs to iterate on data quality based on results.
• Work closely with the other Research Scientists to build shared experimental infrastructure and benchmarks.

Tech stack: LLM post-training — SFT, RL.

Requirements

• Run controlled SFT and RL experiments to measure dataset impact on model performance
• Quantify lift across capabilities including reasoning, tool use, long-horizon tasks, and domain-specific workflows
• Communicate findings with partner labs to drive sales
• Work with internal SPLs to iterate on data quality based on experimental results
• Strong familiarity with LLM training and evaluation methodologies
• Design lightweight experiments and extract actionable insights from messy results
• Work across multiple domains including finance, software engineering, and policy

Green Flags

• Has run controlled post-training experiments end-to-end, can point to a specific data intervention that shifted model behavior in a measurable way
• Comfortable reading messy experimental results, doesn't need clean data to find signal
• Strong quantitative instincts paired with SWE ability, can actually ship the experiment, not just design it
• Has worked adjacent to or inside frontier labs or eval orgs — understands what "high signal data" actually means in practice

Red Flags

• PhD-only researcher profile with no shipping track record, role explicitly prefers pre-PhD builders
• Wants to focus on a single domain — the work spans finance, code, policy, and enterprise workflows

Why Join

• Work directly shapes the datasets leading AI labs use to train next-generation models.
• Base plus profit sharing pushes total cash to ~$250K–$450K, with equity on top.
• Build over theorize — high-leverage experimental work, not a pure-research seat.

Details

• Location: San Francisco, CA
• Work policy: On-site
• Compensation: $150K–$250K base + profit sharing (~$250K–$450K total cash) + equity
• Visa sponsorship: None available
• Employment type: Full-time