Case study
PaellaLab
The arithmetic behind a good rice.
Rice and cooking time in; broth, fat, salt and portions out, from how much water each pan actually loses to evaporation. Around the calculator: recipes, a registry of your own pans, a timer and an AI cooking assistant on Llama 3.3. Installable, offline-capable, bilingual.
A small assistant on cheap inference
Embeddings from Hugging Face, vector search in Postgres, generation with Llama 3.3 70B on Groq, with rate limiting written into the route itself — a hundred messages a day, ten a minute — sized for roughly fifteen users. A frontier model would have cost more per month than the app is worth, and would not have answered a question about rice any better.
One formula, measured per pan
Every pan loses water at its own pace, so each one gets a quick evaporation test. The app turns that measurement into how much broth a given amount of rice needs for the time it will cook, and works out the rest of the quantities from there.
Hydration budgeted island by island
Only authentication and navigation load eagerly; every landing section waits until it is visible. The gated content exists twice on purpose — enforced on the server, softened on the client with a fade — because the first is the rule and the second is the manners.