What’s been the trickiest part of getting AI to work reliably in production for you?

I am just curious about the real work, not the flashy demo stuff.

To be honest, it has not been about picking the right model for us; they are all great in their own way. It has been everything about the model:

• Managing auth and region-specific routing (US vs EU)

• Keeping user context in sync across Redis, Postgres, file storage

• Dealing with random failures that only happen in production, never in dev

• Making sure nothing breaks when we swap one model out for another

Some days it feels like we’re duct-taping 10 systems together just to answer one user request cleanly.

How about you? What’s that one annoying thing that keeps popping up when you’re trying to ship something AI-related? We’re all figuring this out as we go, so thought I’d ask the room. 🤟