All roles

Applied ML Engineer

Up to $350K + equity San Francisco, CA On-site AI model training Seed 67 people

About the role

Who You'll Join


A seed-stage AI startup teaching AI models to judge subjective quality. You'll work on teaching models what great looks like across subjective domains: design, writing, personality, and visual style. The team is small and well-funded, with clear paths to revenue through partnerships with AI labs.


What You'll Do


  • Build the reward models, classifiers and verifiers that judge subjective qualities like design and writing style.
  • Design and run frontier evaluations and benchmarks for domains where human judgment matters.
  • Try new data formats and methods such as RLHF and DPO by post-training open-source models.
  • Work hand in hand with AI labs and creative experts on pilots that prove out new approaches.
  • Own full-stack research pipelines: experiments, training, data, evaluation, and results.
  • Publish blogs and whitepapers on your findings.


Who You Are


  • Two or more years training models in production or research settings, especially post-training and fine-tuning work.
  • Hands-on experience with LLMs and diffusion or multimodal models. Pure classical ML is not enough.
  • Comfortable moving fast through ambiguity. You think like a researcher but ship like an engineer.
  • Excited about the problem itself: you care about quality, design, creativity, and reducing AI slop.
  • Experience at startups or data companies (Scale, Snorkel, etc.). Big tech or purely academic backgrounds rarely adapt here.
  • Bonus: Personal passion for a creative field like writing, design, art, or fashion strengthens your fit.


Tech Stack


Python, PyTorch, LLMs, image models, multimodal models, RLHF, DPO, video models, post-training techniques, data annotation tools.


Why You'll Thrive


  • You join at the inflection point. The team is tiny, the offer includes equity, and you shape the research direction entirely.
  • The problem is intellectually novel. Most AI labs ignore subjective domains. You're solving that.
  • Culture is collaborative and low-ego. Founders came from search and design backgrounds, not cutthroat big tech. A food and dining stipend is included.
  • Work is fast-moving but not toxic. Cycles range from one week for quick post-training runs to quarterly for deeper research, with frequent publishing and customer feedback.
  • Clear growth path, with plans to hire four more roles this year.