<aside> 📍

Role at a glance

• Location: San Francisco, CA. Onsite strongly preferred.
• Commitment: 3 to 6 months. Full-time or 20 hours per week part-time.
• Compensation: $35 to $50 per hour.
• Work Authorization: Astorie can provide standard employer documentation for students who are eligible to obtain CPT authorization through their school. </aside>

About Astorie

At Astorie, we are building the intelligent canvas where creative agents can see, reason, and act throughout the creative process. They proactively explore ideas with you, offer aesthetic inspiration, and learn your taste and way of working over time. Our goal is to build the creative infrastructure that closes the distance between imagination and expression.

About the Role

This role works with Astorie's research engineers to train and evaluate agents and video models for real creative workflows. You should bring depth in at least one track—agent post-training or video model post-training—and will have opportunities to expand into the other as projects require.

You will work with GPU access, well-defined training projects, and close technical collaboration. The role combines rigorous execution with experimental design, analysis, and iteration.

What You'll Do

• Curate, clean, and version agent trajectories, preference data, visual assets, and video datasets for training
• Define dataset schemas, filtering and deduplication rules, quality checks, and clear data documentation
• Run SFT, DPO, RL, distillation, LoRA, or fine-tuning experiments in your primary track
• Define hypotheses, baselines, metrics, and controlled ablations for each experiment
• Analyze how data composition, objectives, and training strategies affect model behavior and creative quality
• Build or improve the rewards, graders, and evaluations needed to measure experimental results
• Maintain reproducible training pipelines, configurations, checkpoints, and experiment logs, and communicate findings and next steps

What We're Looking For

• Evidence of personally running at least one reproducible model training or fine-tuning experiment, with code, configurations, and results you can show and discuss
• Strong PyTorch fundamentals and the ability to modify, train, and debug models and data pipelines
• Experience in at least one track: LLM or agent post-training, diffusion models, or video models
• Familiarity with one or more of SFT, DPO, RL, LoRA, distillation, or preference optimization
• Ability to formulate hypotheses, design controlled experiments and ablations, and interpret noisy results