Applied Scientist, Post-training

Job Description: Teach AI how to reason - safely, transparently, and at scale. How do we move beyond pattern-matching into true machine reasoning? This Applied Scientist role puts you at the centre of that challenge - developing models that can reason, explain their logic, and make verifiable decisions across complex, high-stakes industries. You'll join a well-funded startup building domain-specific reasoning systems and agentic AI for sectors like medtech, aerospace, advanced manufacturing - where reliability and interpretability aren't optional. Your work will focus on post-training large multimodal models, applying the latest techniques in RLHF, DPO, and preference learning to make AI systems more consistent, factual, and aligned with human reasoning. You'll design the frameworks that turn raw model potential into transparent, trustworthy intelligence. You'll develop and optimise post-training pipelines, implement reward modelling for reasoning depth and factual accuracy, and build evaluation frameworks for verifiable, human-aligned behaviour. Working with proprietary and synthetic datasets, you'll run end-to-end experiments and deploy your methods directly into production. You'll bring a background in transformer-based model training (LLM, VLM, MLLM), post-training or alignment (RLHF, DPO, reward modelling), and strong practical skills in Python and PyTorch. Curiosity about reasoning agents, hybrid learning, and interpretability research will help you thrive here. Bonus points for experience in multimodal reasoning, evaluation and verification, or prior research contributions in alignment or reasoning systems. The company has raised $20M+ (Series A announcement imminent) and already partners with Fortune 100 and 500 customers. Founded by an entrepreneur with a prior billion-dollar exit, the AI team alone is scaling from 11 to 40+ this year. Comp: $200K-$320K base (negotiable depending on experience) + bonus + stock + benefits Location: SF Bay Area (remote for now; hybrid later in 2026) If you're excited about defining how AI systems reason, decide, and explain themselves - we'd love to hear from you. All applicants receive a response.