Research Engineer/Scientist

About the company

Our client is an early-stage, research-first AI company in San Francisco, founded by researchers and engineers out of frontier AI labs and top-tier research groups. They build personal, proactive agents that learn how an individual actually works and then pursue outcomes on their behalf, and their research agenda is squarely about understanding human intent and scaling human judgement. The team is tiny and talent-dense, works together in the office, and is still early enough that the research direction is genuinely open.

The role

This is a hands-on frontier research seat — internally a member of technical staff on the research side, somewhere between research engineer and research scientist. You would run experiments and train models for human-judgement scaling, computer use and intent representation learning, and post-train LLM and multimodal agents with reinforcement learning and continual learning. Research bets are owned end to end here: hypothesis, data, training, evaluation and deployment. As one of the first research hires you also help set the agenda rather than execute someone else's. Note that this is the research role, not the client's applied or product AI engineering role.

What you'll do

• Run experiments and train frontier models for human-judgement scaling, computer use and intent representation learning
• Post-train LLM and multimodal agents with reinforcement learning and continual learning
• Work with large-scale screen-recording and behavioural data to model how individual people work
• Own research bets end to end — hypothesis, data, training, evaluation and deployment
• Help set the research agenda as an early member of the team

Nice to have

• Research in agents, computer use, continual learning, or modelling human intent and preferences
• First-author papers at NeurIPS, ICML, ICLR or ACL, or a widely used open-source research repository
• A named model, feature or research team you owned at a large lab
• Distributed training and CUDA experience at scale
• Representation learning or vision-encoder pre-training experience

Tech stack

Python, PyTorch and Transformers; LLMs, reinforcement learning, post-training and continual learning; CUDA and distributed training. The client is explicit that the specific stack matters far less than the research record behind it.

Compensation & logistics

• $300k–$600k base + competitive equity, set per candidate; up to about $650k for researchers coming from a premier frontier lab
• On-site five days a week in San Francisco, with relocation assistance
• Open to candidates who need no sponsorship, to visa transfers and to new visa sponsorships — the client offers full immigration support
• Fresh PhD graduates through senior researchers: the research record counts for far more than years
• Full-time