Every portfolio site dies the same way. You build it, it’s great, and then you ship something worth adding and don’t feel like opening the repo. Six months later the newest thing on it is six months old.
So this rebuild started from a different question: not “what should the site look like” but “what would make updating it free?”
The vault is the source of truth
Everything on this site — every project, the timeline, the résumé, this post — is generated from a markdown vault on my machine that an agent maintains. Projects are files with frontmatter. The résumé is a file. My history is a file. A sync script projects the public-safe subset into this site’s content collections and commits it.
That means the update path for “I shipped something” is a one-line capture, not a code change. The site is a render target, not a place where facts live.
Public by exception, not by default
Most of what I work on is at an employer, which creates an obvious problem: the most impressive work is the least publishable.
The compromise is a confidentiality gate. Employer projects carry two sections — a public-safe summary and internal notes — and only the first is ever eligible to leave the vault. On top of that sits a list of banned terms: internal system names, source systems, incident dates. Every outbound string is checked against it, and a hit aborts the sync rather than warning about it. It’s a small amount of code that lets me talk about the shape of my work without ever leaking its internals.
Fail-closed is the whole design. A publishing pipeline that warns is a publishing pipeline that eventually publishes the wrong thing at 11pm.
Show, don’t tell
The other thing I wanted: stop describing machine learning work and let people run some.
The sketch classifier on the home page is a small CNN — about thirty thousand parameters — trained on Google’s Quick, Draw! dataset, exported to ONNX and executed in your browser on WebAssembly. It’s ~89% accurate on ten classes and the weights are 121 KB, which is smaller than most hero images. Nothing you draw leaves your machine, because there’s no server to send it to.
Constraint worth knowing: GitHub Pages can’t send the COOP/COEP headers that WebAssembly threading needs, so it runs single-threaded. For a model this size that’s irrelevant, which is exactly why the model is this size.
Stack
Astro 5 for the static shell, React only inside the interactive pieces, Tailwind v4, deployed by GitHub Actions. The résumé page doubles as the PDF source — one renderer, so the download can never disagree with the page.
The best part isn’t any of that. It’s that the next time I ship something, I’ll write one line about it, and this site will already know.