About / 2026
Norwin Yu.
Independent AI usage research. I study how models, interfaces and local hardware meet — lately through browser games, agent harnesses and the verification layer that keeps both honest.
01 / Research areas
Research areas.
Six domains, each backed by something shipped, tested or written down. The point is the handoff between them: model behaviour, interface craft, local infrastructure and the writing that makes the experiments legible.
| Domain | Focus | Tools | Selected evidence |
|---|---|---|---|
| 01AI systems & agents智能体系统 | Agent workflows that survive real inputs: retrieval, tool use, memory, routing, guardrails and evaluation. | Agent orchestration · RAG · Evals · Local inference · Prompt systems | Name That LLM; multi-tool agent harnesses; multimodal production pipelines. |
| 02Product experiments产品实验 | Static-first front ends, pragmatic APIs and deploy paths that one person can actually maintain. | TypeScript · React · Node · Python · GitHub Actions | Game + tooling pipeline probes; static deployment; 73-check Playwright regression. |
| 03Spatial interfaces空间界面 | Three.js, WebGL and Godot interfaces where depth and motion make information easier to hold. | Three.js · WebGL · Godot · Shaders · Interaction design | Tactical battlefield; 3D sect mountain gate; cinematic game UI. |
| 04Local infrastructure本地基础设施 | Linux, CUDA, containers and local GPU workflows. Systems I can rebuild from notes. | Linux · CUDA · Docker · Self-hosting · Observability | Two-GPU home lab; local inference; static deploy pipelines. |
| 05Research & writing研究与写作 | Small experiments with methods, caveats and the failed attempts left in the record. | Benchmarks · Data analysis · Visual essays · Documentation | Name That LLM; The 44 GB problem; game build postmortems. |
| 06Interface design界面设计 | Editorial hierarchy, typography and restrained motion, built in code rather than a mockup. | Art direction · Typography · Motion · Prototyping | This site; product surfaces across the games. |
Also in the toolkitGit / GitHub Actions · Figma · SQL / Redis · Docker Compose · headless visual QA · Blender basics.
02 / Principles
Method.
Four defaults that show up in the code, the writing and the way a project gets scoped.
Local before cloud
Private data stays on the desk. Use frontier models where the capability gap is real, not by default.
Ship the loop
A playable slice with an end state beats a polished scene with no game. The fifth run is the test.
Measure, then trust
Prompt-level splits, seed checks and screenshots. A single run is an anecdote, not a result.
Write it down
Method, caveat, failure. The notes are how the next project starts faster.
03 / Now
Currently.
A snapshot, not a résumé. For the full list of work, use the work page; for the workstations and model routing, use Local.
Current experiments
Games, comparisons and coding tools, plus the harness around them. The next step is a tighter loop between generated assets, playtesting, a Blender-assisted comparison and visual verification.
Selected experimentsResearch conversations
Open to conversations around AI usage, game prototypes, developer tools and local inference. The fastest route is a short note with the question, the constraint and the deadline.
GitHub