The hardest technical problem I solved building CaloAI

AI models kept returning macros that didn't add up mathematically. Protein × 4 + Carbs × 4 + Fat × 9 would never equal the reported calories. I ended up building a double-layer verification guardrail that forces every response to be internally consistent before it's served. Small detail — but the kind of thing that breaks user trust silently if you ignore it.