42 things I've built, with what each does. API cost is what the AI coding-agent work on each would have cost at the API prices in effect when it ran; output tokens are what the agents wrote.
Sort by
Show
Matching decompilation for x86-64 C++: checks candidate source against the target binary and says what to change so the compiled bytes match.
A cheap transcription service: upload audio or video and get a transcript from NVIDIA's Parakeet model, with free daily transcripts and an $8 unlimited plan.
A terminal and multiplexer: a fork of WezTerm with fixes for daily mux-server use and AI coding agents treated as first-class panes.
A native YouTube Music player with a music visualizer, written in C on gesso: one 4 MB executable using about 26 MB of RAM.
A small native Twitch app for Linux and Windows: see which channels you follow are live and watch one, with no embedded browser.
A compact, agent-friendly summary of GitHub activity since your last check.
Routes messages between AI coding agents in Wakterm and Telegram or Signal, so agents can talk to each other and to my phone.
A small C library on SDL3 for lean native apps and games that build for Windows, macOS, Linux and the web, with installs measured in megabytes.
Estimates star counts and star history for about 45 million GitHub repositories from public watch events, with issues, pull requests and leaderboards, plus a browser extension that sorts a repository's issues by their authors' reputation.
Benchmarks for the Bend language against OpenMP on CPU and CUDA on GPU.
LLM API prices today or at any past date, so past usage is priced at the rates that applied when it ran.
Shows your real Claude Code rate limits in the statusline, fetched from Anthropic's API.
Field reports from letting AI agents run research campaigns on real projects, and what they got right and wrong.
Token-usage analytics for AI coding tools: reads local Claude Code, Codex, Gemini and other session logs and breaks usage down per project with DuckDB.
One CLI for running bounded, unattended jobs across coding agents (Codex, Claude Code, Cursor, Devin, OpenCode and more), with isolation, deadlines, resumable sessions and recorded results.
A programming language that runs the same code in parallel on CPU and GPU and can prove properties about it.
A sourced encyclopedia of UFO cases, people and claims, built as a static site.
OpenAI's Codex CLI with loops, completion timestamps, tool-free sessions and same-turn failure handlers added.
Searches past Claude Code, Codex and other coding-agent sessions by regex.
Harvest: Massive Encounter (2008) decompiled back to matching C++ and ported to macOS, Linux, Windows and the web.
Autonomous trading research for a market-making competition, with agents proposing and testing strategies against held-out market data.
A fast Zsh distribution with autosuggestions, history search, syntax highlighting and directory jumping built in, measured faster than the usual plugins.
Shared GitHub Actions that publish all my packages, so the release steps change in one place.
Crimsonland (2003) rebuilt twice: a Python reimplementation that matches the original tick by tick, and a matching decompilation that plays in the browser.
A coding font: D2Coding with Nerd Font icons and emoji, the font this site uses.
Plans the cheapest week of meals that meets your nutrition targets, from real grocery prices.
Autonomous Research Bench (ARB)
Autonomous Research Bench, a proposed benchmark that scores the AI research system rather than its answer, by how fast it finds competitive solutions.
Compares U.S. economic results under Democratic and Republican presidents, with permutation tests and false-discovery correction.
Fast, resumable parallel file sync over SSH, with uploads to remote hosts added.
Uploads videos to your own YouTube channel from the command line, with captions, thumbnails and resumable uploads.
Sycophancy Bench: measures whether AI models change their answers under evidence-free pushback, compared with how they respond to real corrections.
A fast YouTube Music desktop client for Windows, Linux and macOS that talks to YouTube's API directly instead of wrapping the website.
A text-first forum with verified eligibility and anonymous posting: the browser proves membership with BBS signatures without revealing the X account behind a post.
A browser extension that improves GitHub's interface.
A fast static site generator in a single binary.
A compact Git-aware Zsh prompt with its own asynchronous Git collector.
A drone flight-control stack: train a policy in a fast simulator, then fly it in the VelociDrone simulator through a real controller path.
A turn-based tactics board game on a hex grid for 2 to 8 players, with a Rust engine in the browser, AI opponents and online play.
A headless client and automation tools for the game Gradient Bang.
Runs many Claude Code workers at once on one codebase, splitting a goal into tasks and producing commits while you sleep.
AlphaZero-style training for m,n,k games like tic-tac-toe in Rust, 75 times faster at self-play than a reference implementation.
Tracks CPU prices against benchmark scores to show which processors are good value.
At the Bank of Canada, I worked across trading, portfolio management, quantitative modelling, and policy before co-leading a $20 billion non-USD credit portfolio. At DELV, Euler, and Veda, I built Python trading and fuzzing infrastructure, worked on market strategy, and analyzed protocol risk using on-chain and internal data. Full resume