Here I explain how I solved what you see on the site and in the Lab: what I decided, what I measured and the code that makes it work. Every post links to the repository's files so you can check them yourself.
Gemma 4 with native function calling, four read-only tools and limits on every run. How the Lab's agent works and how much free quota it spends.
Why the Lab's assistant uses Gemma 4 and not Claude, how it fits in Workers AI's free quota and how it answers only with what the site says.
The demo waited up to 112 s for n8n to wake up on Render. Now the workflow runs on the edge with n8n's rules and a parity test.
Hashes instead of unsafe-inline, Zod without eval, styles as classes, fonts as files, Prism instead of Shiki and a test that fails on any violation.
Statuses, history in the same statement, notes, filters in the URL and a formula-proof CSV. That's how the leads panel became a small CRM.
Every page has its own 1200 × 630 image, drawn at build time with Satori and sharp from its HTML. No runtime cost, and cached between builds.
Every API request opened a TCP and TLS connection to Neon. With the HTTP driver, each query is one request. How I measured, tested and can undo it.
Spanish at the root, English under /en/ with its own paths, translations as a layer over Spanish, and tests that fail if the two versions drift apart.
