Picking the stack

A month in. Most of it has been research, not code. Picking the pieces this thing gets built on, because the wrong choice anywhere blows up either the unit economics or the user experience.

The biggest decision was the AI model. Tested the serious options against the same set of food photos. NZ supermarket meals, takeaway, home cooking, deliberate edge cases. Claude was the most consistent. Not always the highest peak accuracy, but the lowest variance run to run, and the most reliable on output format. When it was wrong it was wrong in believable ways.

For a tool that lives or dies on whether users trust the number it shows them, consistency beat everything else. Claude won for now.

The rest of the stack got the same treatment. Supabase for the database, hosted in Sydney, because NZ data sovereignty matters for a health-adjacent product. Resend for transactional email. Whisper for voice. Vercel for hosting and edge rate limiting.

The thread running through all of it was cost. Every estimate costs money to run. Every email costs money to send. Every database query, every photo stored, every voice clip transcribed. None of it is free, and none of it scales for free either. So every choice got checked against what it would cost at ten users, a hundred, a thousand. Picking the cheapest option every time is a trap. Picking without thinking about it at all is a bigger one.

Now the building can start.