NihilDigit

I enrolled at North China University of Science and Technology in 2023, studying intelligence science.

RAFNet was my first academic project, and also where I learned computer vision by doing it. We built a classifier for student behavior in crowded classrooms on top of GroupRec, an existing hypergraph encoder, with a gated fusion head; I ran the experiments. It's under review at The Visual Computer.

TongXing ran in parallel with that work, but from the deployment side. It won a national second prize in 4C2026. The system supports blind and low-vision users with NPU-based obstacle warning, offline speech, and a local LLM agent for closing touch-first app tasks through voice. The project made the potential of mobile NPUs concrete for me, and it made engineering optimization feel like a field with real room for research and product judgment.

ANVIL came out of the overlap between those two tracks: RAFNet gave me a practical CV foundation, while TongXing made the NPU constraints concrete. Standard flow-based VFI breaks down under INT8 quantization, so I looked for a graph that fit the accelerator instead of forcing the accelerator to imitate a desktop model. The result reuses H.264 decoder motion vectors instead of learning optical flow. The arXiv preprint is public, and I am continuing to reorganize the work into a more systematic version. The project page collects the training code, the Android demo, and the paper artifacts.

waybar-ai-usage, a Waybar widget that shows quota and reset time for Claude Code, Codex CLI, Copilot, and OpenCode Zen, began as a small tool for my own desktop. It became my first real open source maintenance experience after users started filing issues and PRs, including multi-browser support and support for providers they used themselves. It's packaged on AUR.

I also contribute upstream to apps I actually use. As a member of the open-ani organization, I contribute features and issue fixes to Animeko, an open-source anime player — a precious experience in how multimedia systems serve real-world applications, and in building a large-scale open-source project for real. My work spans animeko, mediamp, and anitorrent: player and cross-platform support (Windows ARM64, Linux rendering, touch interaction) and media acquisition (PikPak cloud download, HLS filtering).

A short list of things I'm into: LLMs, Linux, ACGN, and music — rock, folk, and songs that carry history with them. Beyond just using them, I find different language models genuinely fascinating — each one has its own character, and I spend a lot of hours just talking to them.