Applied Research Scientist, Safety and Alignment
San FranciscoFull-timePosted 1mo ago
About the role
Applied Research Scientist focused on robust reward design, telemetry, and production systems to improve alignment of Fleet's reinforcement learning agents. Collaborates with safety researchers to co-design RL environments and tasks for downstream alignment training and evaluation.
Responsibilities
- Reward Design and Telemetry at Production Scale. Study task and verifier design effects on model-alignment incentives; improve reward design and telemetry.
- Off-target behaviors from task assumptions and incentives. Investigate unstated constraints, unsolvable tasks, and rewards that misrepresent success across domains.
- Telemetry and Verifier Generation Pipelines. Build telemetry for observability of undesirable behaviors; create scalable pipelines for verifiers translating measurements into behavioral categories.
- Production Oversight Systems. Develop trace-based analysis to measure agent misbehavior across Fleet evaluations.
- Safety evaluations and benchmark analysis. Develop safety measurements and assess risks in Fleet tasks and external benchmarks.
- Safety benchmarks. Build held-out/adversarial evaluations; audit tasks, data, and graders for contamination, poisoning, ambiguity, and shortcuts.
- External benchmark analysis. Examine benchmarks, datasets, and task designs for unsafe incentives.
- Specialization in high-impact domains. Collaborate with frontier labs on cybersecurity to prevent incidents like the Hugging Face breach.
Qualifications
- Willingness to work directly with data — inspect examples and trajectories, question labels and graders, investigate anomalies, trace results to their cause.
- Strong practical engineering skills to build scalable systems.
- Record of empirical experience in alignment, model behavior, or related field.
- Prior safety fellowships or affiliations (optional) useful but not required.
Skills
Required:Data analysis and trajectory inspectionEngineering for production systemsEmpirical evaluation and model behavior analysis
Last updated Oct 8, 2026. Interested? Apply on the company site.
About Fleet
Fleet - AI-powered workflows
Fleet creates simulated worlds and real-world challenges to better understand and shape the behavior of artificial intelligences.
AI-powered workflowsbessemer-ai-infrastructure-2026reinforcement-learning-platforms
More open roles at Fleet
- Member of Technical Staff, Security Engineer1mo agoSan Francisco, CA
- Founding Recruiter2mo agoHybrid (San Francisco)
- Member of Technical Staff, IT Engineer2mo agoSan Francisco, CA
- Research Scientist, Oversight2mo agoSan Francisco
- Research Scientist, Post-Training2mo agoSan Francisco