Projects
Main work
Peko
A toolchain for shipping one app to macOS, Windows, Linux, iOS, and Android, built around Pekoscript, a statically typed language I designed and wrote the LLVM compiler for.
On holdRust / LLVM 18 / C / Objective-C / Java
Peko Studio, the IDE I built as a Peko app. Pekoscript type errors show up next to TypeScript ones. ProxOS
A bytecode VM, an assembler for its own assembly language, and the start of a kernel, written in C and C++ without AI.
In progressC / C++ / CMake$ prox asm demo.prasm $ prox vm out.prx PROX 5 4 3 2 1 42 Hi!
Assembling and running examples/demo.prasm: a banner, a countdown loop, a subroutine call, and a heap round trip.
Smaller and team projects
Kyro
A JavaScript and TypeScript engine in Rust with a bytecode interpreter and an LLVM ahead-of-time compiler, meant to run inside Peko apps.
I wrote the architecture and spec. Claude wrote the implementation.
PostponedRust / LLVM 18Lark
A systems language that extends C11 with garbage collection, interfaces, generics, and modules, and compiles to C.
I designed the language. Most of the implementation was AI-written.
v0.1.0Rust / Cexport managed struct Person { gc char* name; int age; } export struct Data<T> { T* data_point; } export iface Greet { void say_hi(Self this); // ... }A managed struct, a generic, and an interface, from examples/tour.lark. Collet
An experiment on whether AI coding agents edit Rust better through structured tools (rust-analyzer, tree-sitter, and cargo behind MCP) than through plain file edits.
I designed the experiment. Claude wrote most of the code.
Early resultsRust / Python / MCPArm Semantic Text Shell C1 26% 57% 17% C2 33% 50% 17% C2-forced 6% 65% 30%
Share of tool calls across 11 runs. Even with semantic tools available, the model mostly read and edited text, and never once called the tool that bundles a symbol’s full context. Peko policy checker
Checks an iOS or Android project against App Store and Google Play policy. The lint runs locally, from a CLI, a GitHub Action, or a Claude Code plugin.
RustSidecar
A Mac app I built for my own customer interviews: tracks hypotheses and evidence, manages outreach, and transcribes calls live.
Built with AI coding tools.
Python / Svelte / TauriRitnag
A desktop companion that listens while you work, recognizes repeated questions, and answers them out loud. Built with a team at Cal Hacks 13.0.
Much of the code was AI-assisted.
TypeScript
Loan processing research
Interviews with independent mortgage processors in 2026, and a prototype desktop tool for reconciling loan fees across documents, tested only on synthetic data.
Past researchTypeScript / Electron
The prototype on a synthetic practice loan.