Sweet PalNotes from the people building an app for type 1 diabetes: artificial intelligence, vibe coding and everything in between.
How we build a real AI product: architectures, latency, costs, model evaluation. With the numbers from our own test bench. Home
AI chat cost lives in the tokens a model reads, not the ones it writes. Multiple rounds, prompt caching, one model per role and the real cost per subscriber.
Posted by Simone Tolomelli at · Labels: AI · 6 min read
How we do LLM evaluation before release: reference sentences against the real model, versions switched off but never deleted, and time to first token.
Posted by Simone Tolomelli at · Labels: AI · 7 min read
We took the deciding away from the LLM: JOLs classify in parallel, a script writes the brief, the model only speaks. Sweet Pal now replies in about 1.5 s.
Posted by Simone Tolomelli at · Labels: AI · 7 min read
