Member of Technical Staff - Post-Training Research

NextToken builds specialized models and learning systems for production AI agents, turning the experience of production agents into evaluations, reward signals and post-trained models that improve continuously. Founded by Nitish Kulkarni and Alankar Jain, who led search infrastructure and RL post-training for Google's AI Overviews and AI Mode. $7M raised. As an early technical hire you'll work directly with the founders and see your work deployed in real products.

Who can apply: Visa transfers accepted (OPT, H1B transfer); new sponsorship case by case. On-site 5 days a week in Mountain View, CA (also open to candidates in San Francisco).

What you'll do

• Turn customer workflows and model failures into clear research questions, evaluations and experiments.
• Develop training data, reward signals and verifiers; run SFT, preference optimization and RL experiments.
• Diagnose training failures and use ablations and held-out evaluations to establish what actually improves quality.
• Lead post-training projects end to end, through to a production model or a published result.
• Work with research and systems engineers to deploy models and learn from production behavior.

Requirements

• 2–6 years as a research engineer or applied researcher with substantial hands-on post-training of language models.
• Strong empirical judgment and strong Python and PyTorch skills.
• Experience with agentic RL, distillation, synthetic data, reward modeling or distributed training especially relevant; PhD not required.
• Stack: Python, PyTorch, JAX, vLLM, SGLang, Megatron-LM.
• Visa transfers accepted (OPT, H1B transfer); sponsorship case by case.
• In person 5 days a week in Mountain View, CA.

Compensation: $200K – $350K base plus 0.1% – 2% equity depending on experience; the cash/equity split can be adjusted to suit you. Full-time, on-site in Mountain View.