Agentic RL researcher, post-training (San Francisco or remote) About the company A well-funded voice AI company in San Francisco, building speech models after a recent raise. The research team is growing, with several researcher roles open at once, and is kept deliberately flat. This role reports to the Head of AI. The role You’ll own the RL side of post-training the next text-to-speech model. A big part of how that model gets good is how it’s post-trained, which means building complex, text-based agent loop systems that generate the training signal for RL. You’ll design and run PPO, GRPO, DPO or on-policy loops on model weights, with the environments and rewards that go with them. You’ll also build pipelines that write their own questions and tasks, so training scales without constant manual data curation. Ablations show what moved behaviour, and you’ll work with the speech researchers so what you build in text carries over to audio. What we’re looking for • Hands-on RL or preference learning that trained model weights and changed behaviour in a way you can measure. • You’ve built or led an RL pipeline that scales, ideally one that generates its own questions and tasks (curating an SFT dataset on its own isn’t what this role needs). • You’ve done it at a real company, with real constraints, on chatbot-style or conversational agents. • Bonus: audio or speech post-training, large-scale distributed training, and published or open-source work in LLM post-training. Salary and benefits Up to around $250k depending on experience, plus equity. Healthcare, dental and vision. Logistics Full-time, permanent. San Francisco hybrid or fully remote. Start date ASAP.