Education

B.S. in Computer Science

University of California, Santa Barbara
06/2027

GPA: 3.98

Graduate Coursework: Foundation Models (CS 291A), Reinforcement Learning for Robotics (CS 291K)

Experience

NVIDIA

ML Research Intern, Post-Training | Mentor: Pritam Gundecha
06/2026 – 12/2026 | Santa Clara, CA
  • •Owned the post-training recipe for Nemotron’s PDF knowledge work capabilities, spanning synthetic data generation, SFT, and GRPO; surpassed DeepSeek V4.1 Flash and Qwen 3.8 27B on the GDP.pdf benchmark⁠ (in AA index).
  • •Scaled distributed training across hundreds of GB200/B200 GPUs for single-turn and agentic workloads, using Megatron-LM, NeMo RL, and NeMo Gym to investigate model failures and improve multimodal reasoning.
  • •Open sourced training and synthetic data recipes in NVIDIA NVFlow⁠

    • UCSB AI Group⁠

    Undergraduate Researcher | Advisor: Xin Eric Wang
    10/2025 – Present | Santa Barbara, CA
  • •Led creation of a long-horizon coding benchmark that reframes context rot from static retrieval (needle-in-the-haystack) to agentic performance under realistic user/environment distractors. (COLM 2026 Context Beyond the Window Workshop)
  • •Developed an embodied benchmark for functional Theory of Mind, testing whether agents can act on others beliefs during collaboration rather than merely answer explicit belief questions. (ICML 2026 AI4GOOD Workshop)
    • 03/2026 – 06/2026 | Santa Barbara, CA
  • •Extended ChipAgents coding harness into a long-horizon, multi-agent system with persistent subagents, inter-agent communication, resumable execution, and live subagent steering (TypeScript + Bun).
  • •Built batch execution mode for reproducible, large-scale evaluation of ChipAgents on internal RTL benchmarks
    • 06/2025 – 09/2025 | San Jose, CA
  • •Wrote depthwise convolution kernels in Rust to achieve 16x faster performance on our NPU. Evaluating on MobileNetV2 demonstrated a 5.8x improvement in inference latency.
  • •Explored efficient autoscheduling for CNN/MLP operations by extending Halide to support in-memory compute primitives.
    • Awards
    10/2025

    Won the world's largest collegiate hackathon (3k+ participants, 700+ teams) building personal computer-use agents remotely controlled through FaceTime or iMessage. Released 1 month before OpenClaw.

    Publications

    NVIDIA Nemotron Team (incl. Dylan Lu)

    Dylan Lu*, Surya Gunukula*, Saaket Agashe, Tengxiao Liu, Xin Eric Wang

    Gurusha Juneja*, Dylan Lu*, ..., William Yang Wang, Yali Du, Xin Eric Wang