Nothing here yet.
The jump from URLs and browser history to a navigation stack and app lifecycles is exactly what tripped me up coming from web too. I've found the platform specific bugs only really click once you've shipped something and watched it break on a real device. Was there one Android or iOS quirk that took you longest to figure out?
The conflict resolution part is where I've seen offline-first get underestimated, especially once two devices edit the same record. For a v1 I've found it's often enough to make reads work offline and queue writes with a clear sync status, then add CRDTs only for the screens that truly need collaborative edits. Did you hit any issues with SQLite migrations once real user data was on device?
Routing helps, but a lot of the waste I see is upstream of the model choice. Huge system prompts, full chat history sent every turn, no caching. Trimming that is boring but often cuts more than switching models. Has anyone here seen routing hold up once you count the extra eval work for each model?
Nice use case, building it for an actual user makes it way easier to judge if the AI part helps. Running it locally with Ollama is a good call for a first pass too, no token bill while you figure out the prompt. Did you keep the non AI exam flow as is and add the model on top, or does it change how exams get created?
The second project feeling is real, especially the part where launch messaging and analytics all land at once. I think a small checklist of the two or three things you actually need for a first launch helps more than another tool. Which of those steps do non-technical founders seem to get stuck on first?
The names-not-vibes step is the one people skip. A feature can look dead in the dashboard while two accounts use it every week, and you only find out when they email. Telling those few people before the date, with a clear fallback, seems like the part that decides whether removal goes smoothly. Do you ever keep a feature just because one big account relies on it?
The bit about a coordinator following a specialist's wrong advice is the failure I hit most. Each agent looked fine when I tested it alone, and the final answer was still wrong because of what got passed between them. Grading the whole episode instead of the last message is what finally showed me where it broke. How many reviewed cases did you need before the judge felt calibrated enough to trust?
The invoice total example is exactly how my CI got ignored too. One formatting change from the model and suddenly everyone just reruns until green. Switching to schema checks plus a small normalizer for numbers and dates fixed most of it for me. The part I still struggle with is free text fields. Do you assert anything on those beyond length and required keywords?
The add-on pile-up is the one that got me on a first client build. The base server was cheap, then managed Postgres, backups, logs and an email service each added their own line, and none of it was on the estimate I gave. Now I put every service the app touches in a sheet before picking a host. Did you find egress or the managed add-ons hurt more for small apps?
