Confirmed on the employer's own hiring board on Nov 22, 2026. First seen by Alion on Nov 12, 2026.
Job description
Collision AI is building frontier cybersecurity models to help defend our customers. We’re looking for researchers to develop the training methods and infrastructure behind them, with a focus on post-training and continual learning from real security work.
You’ll take research ideas from an initial hypothesis to deployed models. You’ll build and refine post-training methods, working out why an approach holds up or doesn’t and tracing failures such as unstable training, reward hacking, and loss of exploration. Along the way you’ll identify useful signals in production telemetry, help build training datasets, and develop evaluations that measure how well our models and agents perform. You’ll own the full experimental cycle, combining research with hands-on LLM training.
As an early member of the team, you’ll work directly with the founders and design partners to decide what we build and how. Your work will shape the technical foundations, product direction, and the way we conduct research as a company.
Required qualifications
Experience training large language models.
Proficiency in PyTorch, JAX, or a comparable ML framework, with the engineering skills to turn research ideas into working systems.
Strong experimental judgment: forming testable hypotheses, choosing meaningful measurements, and using results to guide the next experiment.
Bonus qualifications
Experience with distributed training, high-performance computing, or debugging experiments across large compute clusters.
Contributions to open-source ML research, training systems, or evaluation infrastructure.
Evidence of independent technical thinking through research, open-source work, or ambitious personal projects.
A master’s degree or relevant research experience.

