Research Engineer, Training & Inference

About The Role

We are developing reinforcement learning systems at a scale where standard abstractions frequently fail. Unlike labs that operate primarily through high-level wrappers, we own the entirety of our RL stack. This ownership spans from low-level environment simulators and custom communication primitives to our distributed training loops and inference engines.

Key Responsibilities

• Total Stack Ownership: Maintain and optimize our proprietary RL training and serving infrastructure. You have the authority to refactor any layer—from the Python API down to the CUDA kernels—to achieve peak performance for foundation model workloads.
• Optimized Training: maximize the throughput of our reinforcement learning system from data generation to model training with sharded multi-node training and inference algorithms.
• High-Performance Serving: optimize our inference stack for high-throughput reinforcement learning and low-latency LLM production traffic. Tune the inference engine, router, and scheduler, down to custom kernels if need be.
• Compute Optimization: Identify and resolve performance bottlenecks within our distributed clusters, ensuring optimal throughput and memory efficiency for multi-billion parameter models, balancing memory constraints with compute-heavy training cycles.

Minimum Qualifications

• OR BS in Computer Science or a related technical field, or equivalent industry experience
• 2+ years of relevant, hands-on industry experience
• Proficiency in Python
• Experience building or maintaining components within ML frameworks (e.g., PyTorch, JAX, or TensorFlow).
• Proficiency in either:
• Understanding of distributed training concepts and collective communication primitives (e.g., NCCL).
• Practical experience deploying and profiling models on GPU-accelerated cloud infrastructure.

Preferred Qualifications

• MS or PhD in Computer Science, Mathematics, or a related field.
• 5+ years of relevant, hands-on industry experience
• Proficiency in C++
• Experience writing or improving kernels (Triton, CuTeDSL, TileLang, CUDA, CUTLASS, ThunderKittens) to resolve low-level bottlenecks.
• Proven success deploying performant inference at scale using open-source or custom inference engines, routers, etc.
• Direct experience scaling models via FSDP, Tensor Parallelism, or related sharding techniques on multi-node GPU clusters.
• Experience designing reinforcement learning systems for high-throughput training and asynchronous data sampling.

What We Offer

• Unlimited PTO
• 401(k) matching
• 100% employer-paid health, vision, and dental benefits for employees and 50% coverage for dependents.
• HSA available for qualifying health plans

Equal Opportunity Statement

Harmonic is committed to diversity and inclusivity in the workplace. We are an equal opportunity employer and do not discriminate on the basis of race, religion, national origin, gender, sexual orientation, age, veteran status, disability or any other legally protected status.