Careers at Triage
San Francisco · In person
Build institutional alignment.
Triage is an applied AI lab building systems that keep AI agents faithful to an organization's objectives and standards.
We study how behavior changes over the course of a task, how to correct it, and how to make useful autonomy practical.
Explore our work
Runtime judgment and steering
Understand when an agent is drifting from its objective, choose a useful correction, and evaluate what happens next. The challenge spans the full trajectory of a task.
Evaluation and adaptation
Turn reviewed examples and feedback into better judgment. Build evaluations that expose failures, measure regressions, and make changes to behavior inspectable.
Systems for real workflows
Bring the research into the models, routers, and tools organizations already use. Make state, permissions, reliability, and performance hold up through long-running work.
Research methods
- Reward modeling over trajectories
- Learn signals for progress, constraint satisfaction, and recovery across an entire task. Study how process and outcome rewards can capture institutional judgment beyond the quality of a single response.
- Continual learning and retention
- Adapt judgment from reviewed failures and changing institutional requirements. Explore targeted updates and replay of prior examples, using held-out evaluations to measure both new learning and retention of existing capabilities.
- Reinforcement learning for intervention
- Study policy optimization with learned rewards and verifiable outcomes. Investigate credit assignment across multi-step work: when to intervene, which correction helps, and whether the agent recovers and completes the task.
- ML efficiency and adaptive computation
- Explore parameter-efficient adaptation and selective use of expensive inference. Measure how context, latency, and compute budgets affect judgment, and where additional computation changes a decision.
- Counterfactuals and reasoning faithfulness
- Replay decisions under changed instructions, observations, or actions to isolate what drives behavior. Use ablations to test whether a model's stated reasoning predicts what it actually does.
- Adversarial evaluation and transfer
- Generate adversarial trajectories that expose instruction conflicts, objective drift, and failures that emerge over time. Test whether improvements carry over to unseen task families, environments, and models.
How we work
Follow the problem through.
We're a small, technical team working together in San Francisco. Our work connects research, engineering, and deployment: an observation becomes an experiment, an experiment becomes a system, and the system gives us new questions to investigate.
We value clear thinking, direct feedback, and care in execution. Show what you built, explain the decisions behind it, and be ready to change your mind when the evidence does.
Get in touch
Introduce yourself.
If these problems overlap with your work, we'd like to hear from you. Send a short introduction, a project or piece of research you're proud of, and what you'd want to work on at Triage.
info@triage-sec.comRead our Applicant Privacy Notice for how we handle recruiting information.