The Hidden Technical Debt of Managed AI Platforms: A Migration Case Study

Ok I need to vent about something because I'm genuinely pissed off right now.

I tried to migrate Forum BD off Lovable this week. Should've been simple. Point to a new database, deploy on Vercel, done in a weekend.

Instead I spent days untangling garbage I never asked for.

No superuser access to my own database. Couldn't even run a normal dump. Had to rely on some async binary export that only worked with a bleeding edge Postgres version, so I'm sitting there installing beta tooling just to restore MY OWN data.

Then I clean up the codebase, remove their proprietary packages like you're supposed to, and suddenly everything 404s. Turns out they'd been secretly injecting a custom Vite plugin that hijacked the whole build pipeline behind my back. Nobody tells you that. You just find out when it breaks.

And then auth. Google OAuth silently dies on redirect, no errors, nothing. Days of debugging later, I find out they wrapped the entire client in some lazy loaded proxy that initializes just late enough for the token to get stripped from the URL before it's ever read. A race condition, baked in, on purpose, so the "getting started" experience feels smoother.

None of this was my bug. None of it was my architecture. It was decisions made for me, hidden, so I wouldn't leave.

I'm done. I don't have the time or patience to keep fighting a platform that's actively engineered to make leaving painful. Forum BD is paused for now because of this, not because the idea died, but because I burned my energy fighting infrastructure instead of building.

If you're prototyping on these platforms, fine, use them, they're fast. Just know going in that "fast" quietly means "locked in" unless you check every layer yourself.

Lesson learned the hard way.