The AI chose the stack

Picky’s visual layer took less than a day with Stitch; Gemini inside Cline wired the HTML files together ▸ 2:10. Then came the question that used to define senior engineers: what stack do we scale this on? The question went to Gemini ▸ 3:05.

The answer: Next.js, for faster loads and per-restaurant menu updates without redeploying, and PostgreSQL using JSONB, a binary JSON column type that makes queries fast. The honest part, said on camera without embarrassment: “I didn’t even know that existed. I still don’t understand how it works, but supposedly the queries are fast. So it serves me” ▸ 3:42.

Five hours into the migration, the wobble every builder knows: why am I doing this? Why not just plain HTML for the MVP? ▸ 4:33 The question went to GPT, which defended the original recommendation: faster, more maintainable, scales better. “Entonces, yo les hago caso”, so I do what they say ▸ 4:46.

el arquitecto es un consenso de modelos →

Worth examining rather than mocking, because this is how a growing share of architecture decisions now get made. The trust isn’t blind; it’s calibrated differently. He can’t audit JSONB’s internals, but he can audit outcomes: does the menu load fast, does the migration finish, do updates deploy per restaurant. Choosing tools you don’t fully understand on credible advice is also how juniors have always used seniors; the senior just changed species. The risk that remains is the wobble moment, when you can’t tell inertia from wisdom, and the mitigation on display is asking a second model to re-argue the decision from scratch. Two independent yeses beat one; the working demo, filters, cart, taxes and all, beats both…