AI Researcher

The role

You will own the post-training loop by turning real workflow data into models and evaluations. You will improve agent performance through fine-tuning, preference optimization, and building evaluation suites.

What you will do

• Own fine-tuning, preference optimization, and post-training pipelines
• Build eval suites and golden sets to measure changes
• Turn production traces and failure cases into training data
• Work with the platform team on routing, distillation, and cost and latency budgets
• Evaluate fine-tuning versus prompting, retrieval, or scaffolding

Required qualifications

• Strong applied ML background with hands-on post-training experience (SFT, RLHF, DPO, or similar) on LLMs
• At least one model or system you took to real users
• A clear account of how a model or system behaved in production
• Fluency with the modern training stack
• A working grasp of inference economics
• Evaluation discipline

Nice to have

• Published research or open-source work in LLMs and agents
• Agentic evals, tool-use training, or long-context work
• Reinforcement learning environments
• Taking research into production

Benefits

• Real data
• Your work ships
• High-bar team
• Small, early team
• Meaningful equity

Before you apply

• Employment is subject to a standard confidentiality and non-disclosure agreement

About the company

Varick builds AI agents that take over real operational workflows inside the world's largest enterprises. The team includes members from Meta, AWS Bedrock, Citadel Securities, McKinsey, BCG, and Stanford.