I am interested in the systems side of AI: how models are represented, lowered, scheduled, profiled, optimized, and deployed on constrained hardware.

My favorite kind of work sits between research and engineering:

• turning a model idea into something runnable;
• finding where the real bottleneck hides;
• making benchmarks reproducible enough to trust;
• building small teams that can move from uncertainty to working systems.

• Reading systems, ML, architecture, and security papers with an engineer's eye.
• Profiling code until performance numbers become explainable.
• Working close to hardware: CPU backends, mobile devices, heterogeneous compute, memory layout, kernels, and runtime details.
• Writing clean notes, reproducible scripts, and documentation that future-me will not resent.
• Exploring tools that make engineering collaboration sharper, faster, and calmer.

Model systems and inference

• LLM inference pipelines, low-bit model deployment, runtime integration, benchmark design, throughput / latency / memory / power evaluation.

Systems and backend engineering

• C/C++, Python, CUDA, CMake, Linux.
• CPU backend debugging and optimization.
• Familiar with SIMD / NEON / AVX optimization ideas.

Mobile and heterogeneous computing

• Android / iOS deployment workflows.
• OpenCL backend adaptation.
• Mobile GPU kernel optimization exploration.

Performance analysis

• perf, ncu, NVIDIA Nsight.
• CPU/GPU profiling, hotspot diagnosis, kernel-level performance analysis.

Collaboration

• Task decomposition, code review, experiment organization, technical writing, and keeping a small engineering group pointed in the same direction.

• Efficient and trustworthy deployment of large models on edge devices.
• Model safety from a systems perspective: evaluation, runtime behavior, deployment reliability, and hardware-aware constraints.
• Better runtimes for low-bit and non-standard model representations.
• Mobile GPU kernels and heterogeneous scheduling for practical AI inference.
• Reproducible benchmarking infrastructure for model systems.
• The boundary between AI systems, architecture, and cybersecurity evaluation.

I am trying to become the kind of engineer-researcher who can:

understand the model, read the kernel, measure the system, explain the tradeoff, and make the next version faster without making it less trustworthy.

• Email: zybi@stu.pku.edu.cn