Teams at Harvey

Own research and implementation of post-training methods for legal language models, including supervised fine-tuning, reinforcement learning, reward design, verifiers, preference modeling, evaluations, and distributed training. Lead experiments from research direction through production deployment, while debugging training instability, reward hacking, and divergence. Build rigorous evaluation systems and production-quality Python code for long-horizon legal reasoning and agentic workflows.

Build and scale infrastructure for post-training open-weight language models, including distributed training, reinforcement learning rollouts, agent environments, evaluation systems, experiment tracking, and model serving. Own performance, reliability, GPU utilization, checkpointing, failure recovery, observability, and production integration. Partner closely with researchers to turn experimental workflows into dependable systems and accelerate development of legal AI capabilities.

Build synthetic data pipelines, evaluation systems, benchmarks, verifiers, rubrics, and quality metrics for legal AI models and agentic workflows. Partner with researchers, legal experts, and customers to investigate model failures, validate metrics, maintain data quality and provenance, and convert expert feedback and product usage into training signals and improved evaluations.