puddingsworld
Pudding builds the world.
Diving into the digital world — coming up for air in songs and sayings.
I'm GeniusPudding. By day I do large-model research and infrastructure at an AI-chip company. Around that I build agentic apps and terminal tools, along with biosignal AI, surgical vision, audio tools and games.
The long game: a digital world where everything flows, and a bag of tricks for every kind of problem. This site keeps both halves — the projects and the musings.
Now building
updated 2026-10-18 · all projects →
Paperfact
I'm building Paperfact, a desktop reader that turns any paper into a guided reading session: AI commentary paragraph by paragraph, a citation map of what a paper builds on and what built on it, and a reading queue across venues. This week it learned proper math and diagram typesetting, a keyboard-first list view, exportable reading-list bundles, a macOS build, and a round that drafts a presentation-ready slide deck from a close reading.
Mapetite
I'm building a browser extension that curates food places on top of Google Maps. Saved places now sync both ways with Google lists in near real time, candidates are filtered down to actual food, and the first review and visit-tracking cards just landed.
Assistant Hub
I'm building a personal command centre: a task ledger, routines, widgets, a phone view and a desktop sidebar. This week it gained automatic triage of phone captures, a listening player for lecture series, and a weekly watcher for campus talks.
Street KTV & Mandopop pipeline
I'm running a music line from dataset to stage: MandpopDataPipeline turns Mandarin pop into separated stems and time-aligned lyrics, Reprise Studio renders those into abstract lyric videos, and Street KTV is the gig-side app, a performer app on Android with a live audience songbook on this site where listeners scan a QR code and queue songs.
show 6 more ↓show less ↑
TOEFL sprint
I'm on a 16-week TOEFL sprint using my own daily trainer, 90 minutes a day, with a re-diagnosis in early November and a first sitting planned for early next year.
This site
I'm keeping this site as a mount surface for the tools above. The projects page is now discovered from GitHub topics instead of a hand-written list, and the TOEFL trainer picked up a phone listening playlist.
Agent CLI companions
I'm building three small companions for terminal coding agents, Claude Code, Codex and Gemini CLI alike: Kaikou for Chinese voice input, Clip for pasting screenshots where the CLI has no image paste, Listen for spoken replies in my own voice. This week Kaikou cut idle CPU use, picked up automatic mic selection, Simplified-to-Traditional conversion and model warm-up, and gained a bench comparing Whisper, Breeze, Qwen3 and FireRed on my own recordings; Clip finished an audit pass (clipboard never left open, no console flashes, cheaper idle polling); Listen's speech server stopped popping windows, with readback paused until it stays quiet.
Virtual Embryo Challenge
I'm competing in the Virtual Embryo Challenge at NeurIPS 2026. This week's tooling covers pseudobulk-shift baselines, in-silico knockouts, delta and recompose steps for hypothesis building, and submissions straight through the challenge API, with the final deadline in early December.
Cataract surgery video analysis
I'm starting a research line on cataract surgery video. The first study is now scoped around capsulorhexis tracking and surgical-skill assessment, building on a published benchmark.
Xensur Medical
I'm co-building a medical startup. We've placed well in a few startup competitions this autumn and are preparing for a pitch final at the end of October.
Pick one. The page adjusts to who you are.
Tip: this is just a router. Nothing is sent anywhere.
get in touch
Reach me directly.
Email is the fastest way. Mention rough scope, timeline, and budget if it's a project inquiry — that's how I tell you quickly whether I'm a good fit.