JOB-1853
A small, early-stage AI company is developing models that keep learning, aimed at agents running in production, working side by side with customers on real tasks. It needs one of its first technical team members to push post-training research from open question to deployed model.
The work
- Convert customer workflows and observed model failures into targeted hypotheses, test suites and experiments
- Assemble training datasets, reward functions and checkers, then run supervised fine-tuning, preference optimization and reinforcement learning as the problem demands
- Investigate failed training runs, using ablation studies and a reserved test set to find what truly moves quality
- Drive post-training projects from the first question through to either a launched model or a publication
- Hand successful models to research and systems colleagues for deployment and learn from how they behave in production
What you bring
- A solid record of adapting language models after pretraining, with unmistakable personal responsibility for the experiments and what came out of them
- Sharp empirical judgment, including the ability to explain why a method worked and when a benchmark misled you
- Excellent command of Python and PyTorch, enough to implement and troubleshoot the training code on your own
- Comfort with open-ended problems and talking directly to customers and teammates
- A plus if you know agentic RL, model distillation, generated data, reward models or multi-node training; no doctorate needed
On-site in either the South Bay (Silicon Valley) or the San Francisco Bay Area. Visa transfers considered case by case.
