End-to-End AI Product Delivery: Own AI features and products from prototype through production — model selection, serving, evaluation, and iteration — shipping working software rather than research artifacts.
Inference Serving: Deploy and optimize LLM inference across the fleet using vLLM and SGLang. Tune continuous batching, KV-cache management, quantization, speculative decoding, and multi-model routing to maximize throughput and minimize latency and cost per token.
RL & Post-Training Harnesses: Build and operate reinforcement-learning and post-training pipelines using slime (Megatron-LM + SGLang) and Prime Intellect (prime-rl + the Environments Hub / verifiers). This includes reward and verifier design, rollout orchestration, weight synchronization, and keeping long-running training stable.
Evaluation & Iteration: Build eval harnesses and benchmarks that measure quality, throughput, and cost together, and use them to drive fast, data-informed iteration.
Work Across the Stack: Partner with Infrastructure on GPU scheduling and fleet utilization, and with Product on what to build next and why.