I build systems that have to answer to reality: agents that stop before they fabricate, tools that keep humans in the loop, and experiments where the edge cases are the point.
Right now that means AI agents, systems-and-physics modeling, and the awkward but important question of when automation should refuse to act.
cuuper22.pages.dev is my portfolio: a retro×futuristic desktop where the projects run live. ToaruOS-Arnold boots in your browser, Erdos replays a solve, and there's a DOS prompt if you'd rather type. In a hurry? Quick read has everything on one plain page.
| ToaruOS-Arnold · boot it | A windowed x86 desktop OS with its own TCP/IP stack. About 22,000 lines of ArnoldC, where every keyword is an Arnold Schwarzenegger quote. |
| ArnoldC-Native · try it | The compiler that made that possible: a Scala fork with native x86, freestanding C, and kernel backends, plus about 90 new keywords for systems work. |
| Erdos · replay | A Lean 4 Prover/Critic loop that caught its own AI cheating at math. SHA-256 theorem locking, sanitized feedback, 342 tests. |
| gpu_stack · trace a path | GPU training as one SymPy equation graph, from the cost of a token down to the speed of light, with the unknowns left visible. |
| IVC · see the ledger | Indus script research built to make overclaiming structurally hard. One accepted structural finding. Zero readings, on purpose. |
| CurbRun · try it | Android app that finds legal, free curb parking in San Francisco for the next N hours, from 21,268 SFMTA curb rules. |
Also built: PhysicsLab (a phone as a physics instrument, 35 experiments), Manim Director (a coding-agent toolchain for math animation), codex-canmore, Waypoint, claude-sfx, agent-rate-forecast, and the portfolio itself.
- Systems depth:
ToaruOS-Arnold,ArnoldC-Native,gpu_stack. - Evidence discipline:
Erdos,ivc. - Product judgment:
curbrun,PhysicsLab,waypoint-differentiation-lab. - Tooling taste:
Manim-plugin,codex-canmore,claude-sfx,agent-rate-forecast.
The pattern I care about is not "I used framework X." It is whether the project has a real constraint, a clear boundary, and enough engineering taste that someone else can pick it up without decoding my entire brain first.
- Build the thing far enough that the hard boundary shows up.
- Write READMEs as inspection guides, not trophy cases.
Portfolio · Quick read · LinkedIn · Email