Research Scientist - PhD New Grad

Stealth AI Startup; San Francisco (CA); On-site; USD 120,000 - 180,000; Full time

Job summary

Stealth AI Startup in the San Francisco Bay Area is seeking PhD graduates for a full-time, on-site role starting around December 2026 or January 2027. You will work onsite at either San Francisco or Sunnyvale, contributing to the data engine, evaluation stacks, and reliability work essential for frontier AI.

You will help design ultra-realistic environments, verify model performance, and ship research-driven infrastructure that stress-tests agents across domains.

Qualifications

• PhD-level expertise in machine learning, evaluation, and reliability.
• Experience conducting rigorous experiments with reproducible results.
• Strong data-driven approach to research and prototyping.

Responsibilities

• Build agentic environments and simulations for frontier AI.
• Develop programmatic and agentic verification methodologies.
• Close the loop with post-training runs to showcase frontier performance.
• Collaborate with founders and research staff to shape roadmap and advance AI reliability.
• Analyze model failure modes and create scalable data pipelines for test-time compute.

Skills

ML fundamentals

Empirical research

Research mindset

Education

PhD in CS/Math/Physics

Job description

This role is for PhD New Grads who are available to start full time around December 2026/Jan 2027, in person/onsite either San Francisco or Sunnyvale CA.

About the Role

We are looking for Research team members to help us build the data engine for frontier AI. You will develop the high-fidelity environments and evaluation stacks that the world’s leading AI labs rely on to stress-test their most advanced agents. You will work across domains including Computer Use, Enterprise MCP/Toolcalling and Coding. Verifier Design, Simulated Personas, Benchmarking Personal AGI are some of the research areas we work on.

Responsibilities:

• Build Agentic Environments: Design and implement the next generation of "SimLabs", ultra-realistic, long-horizon simulation environments where agents learn to navigate ambiguity and maintain context.
• Programmatic and Agentic Verification: Develop rigorous, policy-aware judges and evaluations that measure genuine capability and safety beyond simple benchmarks.
• Close the Loop: Design and execute high-quality post-training runs (CPT, SFT, RL) to deliver frontier performance on open-source models using curated, high-signal data.
• Collaborate: Work daily with the founders and research staff to shape the roadmap and push the state-of-the- art in AI reliability.
• Work on analyzing model failure modes and creating frontier data pipelines which scale with test-time compute.

Must Have's

We are looking for individuals who demonstrate a rare combination of technical depth, research intuition, and high agency.

• Technical Foundation: PhD in a technical field (CS, Math, Physics, etc.)
• ML Fluency: A principled understanding of foundation models, including how they are constructed, evaluated, and optimized.
• Empirical Mindset: Experience conducting research or technical experiments with a focus on reproducibility and data-driven results.

Nice to Haves

• Research Taste: You have a strong intuition for identifying what matters in complex problem spaces. You can balance deep research exploration with the pragmatism needed to ship a product.
• Impact-Driven Agency: You care about outcomes, not just activity. You don't wait for a ticket; you identify gaps in the system, build the solution, and ensure it moves real-world metrics for frontier AI labs.
• Domain Expertise: Prior experience with Reinforcement Learning (RLHF/RLAIF), simulation systems, or building long-horizon agentic environments.
• Proven Track Record: A history of contributing to influential ML research (e.g., publications at NeurIPS, ICLR, ICML) or maintaining high-impact open-source projects.
• Post-Training Experience: Experience fine-tuning or evaluating large-scale models to deliver "frontier performance" on open-source benchmarks.
• Engineering Rigor: A solid foundation in software engineering with the ability to build robust, scalable infrastructure. You should be comfortable in a Python-friendly, CLI-first development environment.