AI SYSTEMS · EFFICIENT INFERENCE · EMBODIED AI
JACK HE
I build efficient AI systems, from inference kernels and on-device models to embodied intelligence.
AI Kernel Engineer Intern @ Quadric
UCLA M.S. Computer Science · Expected December 2026
01 // ABOUT
From models to the real world.
I’m Jack He (Zhe He), a Computer Science master’s student at UCLA. My work connects machine learning with the systems that make it run efficiently.
At Quadric, I develop and optimize inference kernels for the Chimera GPNPU. Previously, I worked on humanoid manipulation and urban navigation in Bolei Zhou’s lab, and generative-model memorization in Cho-Jui Hsieh’s lab.
My Chinese name is 何喆 (Hé Zhé). Outside of research and engineering, I enjoy snowboarding and golf and have taught generative AI at UCLA.
3.5×
Linear-attention decode speedup at Quadric
7.6×
Matmul efficiency improvement at Quadric
2.05×
On-device VLM speedup over project baseline
02 // EXPERIENCE
Where I’ve made an impact.
Quadric
AI Kernel Engineer Intern
Implemented Qwen3.6 linear-attention kernels from scratch for prefill and decode on the Chimera GPNPU, supporting production inference alongside MoE blocks. Optimized linear-attention prefill/decode and hardware-level routing, achieving a 3.5× decode speedup. Optimized matmul and GEMM kernels for the GPNPU, achieving a 7.6× matmul efficiency improvement. Improved FX64 square-root and division implementations, resolving precision-scaling issues to improve numerical accuracy in edge cases.
Chimera GPNPULinear attentionGEMMFX64
UCLA · Bolei Zhou Lab
Research Assistant, Embodied AI
Developed a vision-language-action (VLA) pipeline for humanoid manipulation, combining VLM-based instruction following and imitation learning with demonstrations collected through VR teleoperation. Automated the URBAN-SIM asset pipeline with GPT-4o, Grounded DINO, and SAM, supporting 10,000+ interactive 3D assets for procedural obstacle placement. Led sim-to-real deployment on COCO robots, evaluating navigation policies across 50+ unseen urban scenarios.
VLAImitation learningVR teleoperationSim-to-real
UCLA · Computational ML Lab
Research Assistant, Generative Models
Developed a training-free fingerprinting method using layer-wise memorization profiles; studied embedding-space selection for diffusion and GAN models with ViT and CNN encoders.
Diffusion modelsGANsViTCNN
03 // PROJECTS
Selected builds & experiments.
On-Device VLM Inference Optimization ↗
2.05× faster inference · Oct–Dec 2025
Optimized Qwen3-VL-2B inference on Snapdragon 8 Elite, achieving a 2.05× speedup over the project baseline. Configured llama.cpp with importance-matrix-based Q4_0 quantization, memory locking (mlock), and CPU affinity for on-device inference. Fine-tuned with LoRA for visual question answering, evaluating both task quality and inference performance.
llama.cppQ4_0LoRASnapdragon 8 Elite
High-Performance C++ Web Server ↗
1,000+ concurrent users · Mar–Jun 2025
Built an asynchronous C++ web server using Boost.Asio, Redis caching, and PostgreSQL connection pooling to support 1,000+ concurrent users. Led a four-person team using TDD; automated CI/CD with Docker and Google Cloud Build.
C++Boost.AsioRedisPostgreSQLDocker
EEG (Electroencephalography) Signal Classification ↗
Machine Learning
Explored a variety of model architectures for EEG signal analysis, including CNN, RNN attention-based models, Transformers, and hybrid models.
Google TripBud ↗
Software Development
Software Product Sprint project, a trip budget planning web app
Fairness and Factuality of LLM ↗
Machine Learning
Embarked on evaluating and enhancing the performance of Large Language Models (LLMs) in detecting fairness and factuality in textual claims.
Bruin O Bruin ↗
Software Development
Lead a team of 5 to create Bruin ’O’ Bruin, a card-elimination game web app
Text Guided Image Editing (DiffEdit) ↗
Machine Learning
Leveraged DiffEdit, an innovative text-conditioned diffusion model for semantic image editing, integrating it with BLIP and other models into an interactive framework for a seamless end-to-end image generation and editing pipeline. Additionally, pioneered a novel technique for text-guided mask generation in DiffEdit, enabling precise object segmentation through textual queries.
Super Peach Sister ↗
Software Development
A 2D game where players control Princess Peach to save Mario
04 // RESEARCH
Ideas, experiments & publications.
arXiv · 2024
Layer Choice for Memorization Detection and Fingerprinting for Generative Models
Jack He, Jianxing Zhao, Andrew Bai, Cho-Jui Hsieh
05 // TOOLKIT
Tools behind the work.
Languages
C++CPythonSQLShellJavaJavaScript
ML & Inference
PyTorchTensorFlowNumPyscikit-learnllama.cpplow-bit quantizationLoRA fine-tuning
Kernel Optimization
Linear attentionprefill/decodematmul/GEMMFX64 numerical functions
Systems & Tools
Boost.AsioRedisPostgreSQLGitDockerGoogle Cloud PlatformGoogle Cloud Build
06 // EDUCATION
A foundation at UCLA.
M.S. in Computer Science
Expected Dec 2026
University of California, Los Angeles (UCLA)
B.S. in Computer Science; Double Major in Applied Mathematics
Jun 2025
University of California, Los Angeles (UCLA)
- GPA: 3.92/4.0 (undergraduate)
- Dean's Honors List, 2021–2025
- Relevant coursework: Efficient Deep Learning, Advanced Deep Learning, Machine Learning, Computer Vision, Natural Language Processing, Reinforcement Learning, Algorithms, Linear Algebra, Probability & Statistics
07 // CONTACT
Let’s talk AI & systems.
Interested in efficient inference, embodied AI, or research collaboration? The best way to reach me is by email.