Staff AI Engineer — On-Device Models & Training

Ghost; San Francisco (CA); On-site; USD 180,000 - 240,000; Full time

Job summary

Ghost is seeking a specialist to advance post-training language models that run locally on Core hardware. You will post-train models, develop reinforcement learning methods, and shape on-device agent behavior with an emphasis on latency and privacy.

You should have strong Python and PyTorch skills, with experience building training pipelines, datasets, and evaluations, and be comfortable working on-site in San Francisco with cross-functional teams.

Qualifications

• Experience post-training language models with supervised fine-tuning, reinforcement learning, or preference optimization.
• Strong Python and PyTorch skills, with experience building and debugging training pipelines.
• Experience creating datasets and evaluations that produce measurable improvements in model behavior.
• Familiarity with tool-using agents, multi-step environments, and the challenges of training against task outcomes.
• Strong experimental judgment: you can diagnose failures, test hypotheses, and distinguish benchmark gains from meaningful product improvements.
• Ability to work on-site in San Francisco.

Responsibilities

• Post-train open models to power personal agents that run locally on Core.
• Improve tool use, reasoning, and reliable execution across complex, multi-step tasks.
• Build training datasets and feedback pipelines grounded in realistic agent workflows.
• Develop supervised fine-tuning and reinforcement learning methods that translate model capability into useful product behavior.
• Build evaluations that measure task completion, personalization, and when agents should act or ask for guidance.
• Work closely with the agent and inference teams to improve quality within on-device memory, compute, and latency constraints.

Skills

Python

PyTorch

On-device inference

Tools

PyTorch

Job description

Ghost is seeking a specialist to advance post-training language models that run locally on Core hardware. You will post-train models, develop reinforcement learning methods, and shape on-device agent behavior with an emphasis on latency and privacy.

You should have strong Python and PyTorch skills, with experience building training pipelines, datasets, and evaluations, and be comfortable working on-site in San Francisco with cross-functional teams.