Akoya

Each generation — the day-by-day details, the itinerary summary, the tips — has its own Jinja2 template rendered with the trip's parameters. Prompts became files you could diff, review and change without touching request handling.

This is unremarkable practice now. In early 2023 the common alternative was f-strings scattered through view code.

A full itinerary takes long enough that a spinner reads as broken. The details endpoint streams the completion to the browser over server-sent events, so text appears as it is generated. There was no framework helper for this; it is a streaming HTTP response with the content type set and caching turned off.

The AI endpoints carry their own throttle class, separate from the rest of the API. A generation endpoint is a public URL that costs real money every time someone loads it, and rate limiting it is the difference between a product and a donation.

This is the same argument the studio would make again years later about metering agent costs — it just arrived here first.

Every call is deterministic. A travel plan that reshuffles itself when the user refreshes is not a plan, and creative variance is not a feature when someone is deciding what to do on Tuesday.

Temperature zero throughout. The model was there to organise information the user had supplied, not to be interesting.

Throttling the generation endpoints was part of building them. It is the cheapest possible version of the metering argument we would make about agent systems years later, and the reason it is worth pointing at now is that the instinct is older than the vocabulary.

Templating them made the product's behaviour reviewable. When the output was wrong, the fix was a diff on a template rather than an archaeology expedition through a view function.