I build foundational systems from scratch — an HTTP stack on raw TCP, Git from first principles, parsers, flight simulation — and then do the part most from-scratch projects skip: prove they're correct. Differential suites against the reference implementation, wire-level conformance tests, benchmark harnesses that live in the repo next to the code they measure.
Not vanity metrics — each gauge is a number produced by a test or benchmark harness committed in the repo it describes. The panel itself is generated by a script, because a hand-drawn precision instrument would be a contradiction.
One channel per repo, redrawn daily by a workflow from the commit stats API. Every channel is auto-gained to its own peak — and the peak is printed, so the gain is never a lie.
Every repo ships a live, interactive playground on its own Pages — same night, four rooms. Click through; each one starts flying itself until you touch it.
GitEngine · review room — add, commit, branch, merge; every SHA verifiable against native git
|
PulseHTTP · strip chart — starve the token bucket yourself; 401 ≠ 403, live tallies
|
JSON-LP · railroad — the repo's real parser via Pyodide; every keystroke is a differential test
|
wc-Go · counting room — watch a rune get cut by a 32KB seam and counted correctly anyway
|
| What it is | The proof | |
|---|---|---|
| GitEngine | Git rebuilt from first principles in Go — 17 commands, three-way merge (diff3), rebase | Byte-for-byte SHA parity with native git, enforced by a differential test suite; 2× git's bulk ingest — watch hashes cascade live |
| PulseHTTP | Production-shaped HTTP/1.1 stack on raw TCP in Go, zero dependencies — router, auth, rate limiting, LRU+ETag cache, TLS, keep-alive proxy | 179K req/s @ p99 2.2 ms by its own harness; 39-test conformance suite — fly the pipeline yourself |
| JSON-Lexer-Parser | JSON lexer + recursive-descent parser in pure Python, typed errors, bounded depth | 98-case differential vs RFC-strict stdlib: 0 disagreements; 572K tokens/s — run it in your browser |
| groundschool-skill | A Claude Code skill that reads a repository and writes you a course about it — local-first dashboard, verified code anchors, append-only updates | Self-hosting: the live demo is the skill run on its own repo |
| wc-Go | Concurrent, streaming wc in Go — 32KB chunks, goroutine-per-file |
59/59 differential vs coreutils; lines 3.9×, chars up to 12.8×, 100-file fan-out 3.5× — watch the boundary carry |
- Differential testing over unit vibes. When a reference implementation exists — native git, the RFC-strict stdlib parser — the test is agree with it on everything, not pass the cases I thought of.
- Measured, not estimated. A performance claim without a committed harness is a rumor. Every number above has a script you can run.
- From scratch, on purpose. Rebuilding a thing is how I earn the right to reason about it — the wire format, the index layout, the state machine — not a bet against using the real one in production.
Flight-simulation engineering — Unity/C#, multiplayer VR training systems, ML-assisted flight-data analysis. The night-cockpit look above is an occupational hazard.
READ THE CODE · PROVE IT BY CHANGING ONE THING · WRITE IT DOWN