Description
At Harvey, we're transforming how legal and professional services operate by combining frontier agentic AI, an enterprise-grade platform, and deep domain expertise.
As a Research Engineer, Infrastructure in Harvey Research, you'll build the systems that power how we adapt and improve leading open-weight models for legal work.
Responsibilities
- Scale our post-training stack, supporting SFT and RL on leading open-weight models.
- Build distributed rollout infrastructure that coordinates model inference, tool execution, and collection of training experience for long-horizon tasks.
- Build the infrastructure for reproducible agent environments where models can interact with documents, tools, and external systems during training and evaluation.
- Own the performance and reliability of training runs, including GPU utilization, checkpointing, failure recovery, and observability.
- Identify and remove bottlenecks across training and inference to shorten the path from research idea to useful result.
- Build the internal tools and services the team runs on: training data management and versioning, running evals and browsing results, and launching and tracking training jobs and experiments.
- Integrate models into production, including serving, latency, cost, and online evaluation, so research reaches customers quickly and safely.
- Partner closely with researchers to turn experimental workflows into dependable systems while keeping room for rapid experimentation.
Requirements
- Experience building ML infrastructure, distributed systems, or research tooling, with ownership of performance, reliability, and operation.
- Strong Python and generalist software engineering skills across backend services, data systems, and internal tools, with familiarity with PyTorch, JAX, or similar ML frameworks.
- Experience with distributed training or inference, including diagnosing performance bottlenecks and failures across the stack.
- Understanding of post-training workflows; experience with RL rollouts, agent environments, or sandboxed tool execution is a plus.
- Practical judgment about what to build versus adopt, and the ability to partner with researchers and ship quickly as needs evolve.