pretrain

  • Groq

    08/25 → 12/25

    • Helped ship new load-balancing algorithms for Groq's datacenters.
    • Brought up a Kimi model end to end, from Hugging Face to inference execution on LPUs.
  • NVIDIA

    01/26 → 08/26

    • Won first place in the annual intern competition (software category), where I built a low-latency distributed inference engine for robotics models, running on multiple racks of LPU clusters.
    • Built my own performance engineering tools and shipped new compiler features for performance analysis.
    • Dove into compiler theory and wrote my own small loop-optimization pass for the LPU compiler.