I enjoy working close to the metal: designing high-performance networking and Linux tooling, exploring GPU compute and inference, and learning by building media and image-processing systems from first principles.
⚡ My favorite problems sit at the boundary between software, hardware, and the network.
| Project | What it does |
|---|---|
| HAMi-core | Transparent, in-container GPU resource control for CUDA workloads. |
| vkop | A simple AI inference engine based on Vulkan. |
| vkcompbench | GPU compute capability benchmarking with Vulkan. |
| ffmime | Finds MIME codecs for MSE in WebAssembly. |
| tsanalyze | MPEG-TS parser built for learning and experimentation. |
| ffpic | Hands-on image decoding for BMP, PNG, JPEG, WebP, and HEIF. |
| pktperf | Linux pktgen extensions and helper scripts for packet testing. |
| pycbpf | Compiles classic BPF into C code for BCC workflows. |
| j2depkt | A DSL inspired by Scapy for packet capture and DPDK RTE_FLOW. |
| feishu-bot-webhook-action | GitHub Action that forwards repository events to a Feishu bot. |
| private2fa | Open-source TOTP authenticator vault. |
| arxiv_feishu_bot_action | Collects arXiv papers and sends personalized Feishu updates. |
Linux & networking eBPF • XDP • DPDK • packet capture • load balancing
GPU & acceleration Vulkan • OpenCL • CUDA resource control • inference
Media & vision MPEG-TS • codecs • image formats • WebAssembly
Developer tooling TypeScript • Python • Go • GitHub Actions
Learning by building parsers • DSLs • benchmarks • protocol tooling
There is much more in the workshop—from OpenCL-Benchmark, malitop, OrinSwitchCli, and many more network/system experiments.
Profile metrics are temporarily hidden to avoid external service dependencies.
Build it. Measure it. Understand it. Make it faster.