Education

B.S. in Computer Science

University of California, Santa Barbara
06/2027

GPA: 4.0

Relevant Courses: Data Structures & Algorithms, Machine Learning, Generative AI/LLMs, Distributed Systems

Experience
10/2025 – Present | Santa Barbara, CA
  • •Building novel embodied AI collaboration benchmark with NVIDIA Isaac Sim and Meta Habitat
  • •Leading independent research project exploring theory of mind capabilities in multimodal SWE agents
  • •Made depthwise convolutions 16x faster by optimizing a Rust-based ML compiler for in-memory compute devices
  • •Reduced compile time of tinyML models from 40 mins to 2 secs—then achieved 5.8x faster inference for MobileNetV2
  • •Built GitLab CI/CD pipeline to mathematically verify compiled NPU instructions against ONNX and PyTorch
  • •Designed a custom Halide C++ autoscheduler that jointly optimizes tensor fusion, tiling, and memory placement
  • •Developed a scalable C-based QA pipeline that accelerated IoT evaluation kit testing by 3× during bring-up
  • •Built Python benchmarking tools to measure and profile energy efficiency on customer prototypes (Google and Tile)
    • Projects & Awards
  • •Engineered a computer-use AI agent enabling complete control over personal Macs entirely through FaceTime
  • •Integrated speech generation and understanding in a Flask backend for real-time speech interaction in the FaceTime call
  • •Developed an React + Electron interface to locally prompt the Mac-use AI agent and visualize reasoning steps
  • •Trained a Mixture-of-Experts RL agent achieving >90% human-level completion rate on Geometry Dash levels
  • •Optimized OpenAI Gym RL pipeline to achieve 2× faster convergence in DQN, PPO, and GRPO training
  • •Built a C++ mod for real-time 60 FPS TCP exchange of screenshots and actions between the AI agent and game
  • •Built an on-device voice LLM agent with <1 s speech latency while maintaining real-time interaction
  • •Optimized tool-augmented reasoning with LangChain to achieve +12 point gain in answer quality (LLM judge)
  • •Enhanced RAG pipelines with GraphRAG to improve relevance by ~25%

    • Baddy Buddy AI Coach | SB Hacks 1st Place Entertainment Track

    Winter 2025
  • •Applied ViT models to detect shuttlecock + player position with 94% tracking precision in badminton videos
  • •Built a Next.js + Flask full-stack system integrating Claude 3.5 Sonnet as an AI coach for fine-grained technique analysis
  • •Deployed for 15+ UCSB club athletes, improving training efficiency through motion-pattern feedback and rally summaries
    • Techincal Skills
    Languages — Python, C++, JavaScript, Rust
    Frameworks — Flask, Node.js, Next.js, React, TailwindCSS, Unity XR
    Machine Learning — PyTorch, Transformers, ONNX, OpenCV, LangChain, NumPy, Halide