Harmonic
Research Engineer, Training & Inference
About this role
Harmonic seeks a Research Engineer to optimize and maintain their custom reinforcement learning and inference stack, from Python APIs to CUDA kernels. You'll focus on maximizing throughput for training and serving high-performance AI systems that solve complex mathematical reasoning problems.
What you'll do
- Maintain and optimize proprietary RL training and inference infrastructure across all layers
- Maximize reinforcement learning throughput through sharded multi-node training and inference optimization
- Optimize inference serving stack for high-throughput RL and low-latency LLM production traffic
- Identify and resolve performance bottlenecks in distributed clusters for multi-billion parameter models
- Write or improve custom kernels to resolve low-level compute bottlenecks
- Design systems balancing memory efficiency with compute-intensive training cycles
What they're looking for
- Python proficiency
- ML framework experience (PyTorch, JAX, or TensorFlow)
- Distributed training concepts and collective communication (NCCL)
- GPU-accelerated infrastructure deployment and profiling
- C++ (preferred)
- Custom kernel development (Triton, CUDA, CUTLASS)
- Multi-node GPU cluster scaling (FSDP, Tensor Parallelism)
- Reinforcement learning system design
Benefits
- Unlimited PTO
- 401(k) matching
- 100% employer-paid health, vision, and dental for employees; 50% for dependents
- Health Savings Account (HSA) available
- Work on cutting-edge mathematical reasoning AI
- Elite technical team environment
Opens the application — the Jobs AI extension fills it for you. Set up autofill
Opens the official application on the employer’s site. No login required.
Harmonic
Harmonic builds AI systems for mathematical reasoning and theorem proving, combining reinforcement learning with formal methods to achieve rigorous, verifiable problem-solving. The company is hiring Research Engineers to optimize its ML infrastructure and algorithms, Software Engineers to productionize research into scalable systems, and full-stack engineers to build user-facing products around these capabilities.
View all jobs at HarmonicLikely interview questions
- Walk us through your experience optimizing ML training or inference pipelines. What was the biggest bottleneck you identified, and how did you resolve it?
- Describe your hands-on experience with distributed training frameworks like PyTorch's FSDP or tensor parallelism. How have you debugged performance issues across multi-node GPU clusters?