R&D at ByteDance CapCut · Building practical AI systems — measuring first, optimizing second.
I am an R&D engineer on the CapCut team at ByteDance, with a personal focus on AI infrastructure and models. I learn by shipping small systems, inspecting their bottlenecks, and keeping experiments reproducible.
My public work is deliberately systems-oriented: LLM inference, serving efficiency, model behavior, and the engineering that turns a capable model into a dependable product. This profile contains no internal product, customer, or company implementation details.
observe → benchmark → profile → optimize → document → repeat
- Studying vLLM from first principles: PagedAttention, KV cache management, continuous batching, and scheduling.
- Running controlled inference experiments on Kaggle T4 × 2.
- Comparing throughput, TTFT, TPOT, tail latency, and memory efficiency — not just one headline number.
- Building stronger intuitions for how models, kernels, memory, schedulers, and user experience connect.
Python · Go · Java · CUDA · PyTorch · vLLM · LLM Serving · Docker · GitHub Actions
| Project | What it is |
|---|---|
| vLLM T4 Lab | Reproducible vLLM experiments and system notes for Kaggle T4 GPUs. |
| JEV RSI Lab | Policy-gated Memory-RSI experiments on public and synthetic data. |
| AI Systems Notes | A public notebook for experiments, benchmarks, and performance notes. |
A fast demo is nice. A measured system that explains why it is fast is better.
This profile is the public front door for ongoing experiments. Each project should answer three questions:
- What constraint are we working under?
- What did we measure?
- What changed, and what did it cost?
Signal over noise. Systems over slogans. Curiosity over cargo cults.