The example projects aren't just demos — they're evidence of a way of working. When an AI writes and maintains software, the good patterns change. These are the ones that show up again and again across the projects, each grounded in real code.
Not slideware. Every claim below points at something concrete in the repositories.
Across ~20 projects in the wider workspace. FeatureBoard — the platform that builds the rest — is itself the biggest example of every paradigm below.
Every time data crosses a boundary — a file loaded, a surface parsed, a mesh repaired — the code says out loud where it came from, what it tried to do, whether it worked, and if not, why. When an AI is the one debugging, dense, tagged, structured logs are worth more than clever code. You can always turn verbosity down; you can't recover a reason that was never printed.
The tessellator narrates every face: [STEP-EDGE], [REPAIR], [SNAP-WELD], [PLANE-TESS]. A single run emits a full provenance trail of the geometry decisions.
When something isn't in the right shape, the instinct isn't to print "please run X and come back." It's to do X. If an AI response comes back badly framed, the framing correction lives inside the pipeline and re-runs automatically. If a mesh isn't watertight, it gets repaired before the solver ever sees it. The user is offered the outcome, not homework.
InfinitusPie routes review-response generation through OpenClaw and auto-fetches three drafts the moment a review lands — no "click to generate." STL Wind Tunnel runs STLRepair to watertight the mesh automatically as a pipeline stage, not a manual pre-step.
Heavy work runs as explicit pipelines you can inspect and re-run; light-but-private work runs at the edge, on the user's own device. The split is deliberate: train and process on the workstation, infer where the data already lives.
Gaussian-splat training and the GPU video engine are staged pipelines (COLMAP/nerfstudio, FFmpeg/NVENC, YOLO). SlopRadar does the opposite end of the spectrum: it runs Gemini Nano in the browser — the feed never leaves the device.
A surprising number of "hard" features become easy once you reframe them as: give the model context, ask for a judgement, get structured JSON back. The engineering is in the framing and the schema, not a pile of heuristics.
MarriageBank's "AI Banker" takes a plain-English activity ("back rub") and returns a scored deposit/penalty — a relationship-effort ledger built entirely on model judgement. SlopRadar hands a post to Gemini Nano and gets back a JSON slop/not-slop classification. Same shape, very different domains.
When a machine writes the code, "does it look right?" is the wrong question. Every feature and every bug fix ships with an executable check that proves the behaviour — so correctness is demonstrated, not eyeballed. The test suite is the review.
1,420 executable test specs across the projects — one per feature (test_*f-*.js) and per bug (test_*b-*.js), traceable to the item that motivated it. FeatureBoard: 815 · CADSolver: 106 · InfinitusPie: 102 · STLWindTunnel: 94 · Letter Launcher: 83 · PointCloud: 39 …
FeatureBoard also runs a nightly suite — named checks like "Project Dashboard Load Test" and "API Health" — with 295 nightly test reports on disk. Correctness is demonstrated on a schedule, not eyeballed once.
Features don't start in an editor — they start as structured metadata. A living featurelist.md, buglist.md, a project MAP.md, and an agent_work_log.md give the AI durable memory: what's built, what's broken, what changed, at what cost. That metadata is what lets features be planned, tracked, and empowered automatically.
FeatureBoard tracks 452 shipped features and 431 fixed bugs as structured entries in a 208 KB featurelist.md + buglist.md, backed by a 480 KB agent_work_log.md (1,457 entries) and 1.8 MB of behavioural analytics. Each work-log line carries the task ID, files touched, adds/deletes, and even token cost — e.g. Task: MBF-22 · Files: marriagebank.{html,css,js} · Add: 199 Del: 100. Plans in, provenance out.
Where most projects would npm install someone else's engine and glue it together, these implement the hard part themselves. Not out of stubbornness — because owning the core means it's understood, debuggable, and bendable to the exact problem. The imports are for the boring parts; the interesting part is hand-written.
Gaussian Splats is our own WebGL splat renderer — the PLY parser, the spherical-harmonic evaluation, and the view-sorted gaussian-quad rasterizer are written from scratch, not gsplat.js or a ported viewer. The CAD Solver hand-writes its whole solid modeling kernel — STEP (ISO 10303) parser, surface tessellator, and B-rep modeling ops up to variable-width edge blends — instead of pulling in opencascade.js. The Wind Tunnel is a from-scratch flow solver, volume mesher, and free-surface water model — not OpenFOAM. The Point Cloud Visualizer runs a custom wavelength-propagation engine. Comet Rex has its own THREE-free world model. And the Rust Installer takes it to the extreme: standard library only, zero crates.