Baby Meals

Personal project · AI meal suggester for babies starting solids

Meal suggestions for babies starting solids, fitted to a family's own kitchen - dal, ragi, and last night's sabzi, not rice cereal and jarred purées. It suggests a next meal, tracks what my daughter actually ate, and folds her reactions back into what it suggests next.

Before I wrote the app, I ran a small evaluation to answer a smaller question first: does how you ask an AI for meal ideas actually change what you get back? Getting this wrong isn't a bad suggestion - it's an allergen or a choking hazard, handed to a baby.

The eval, in short

The results were not what I expected going in.

1. The prompt barely mattered. Three genuinely different prompts - plain, few-shot, and a safety checklist - scored 4.16, 4.21 and 4.22 out of 5. My grader disagreed with itself by up to 0.23 when scoring the same output twice - a wider margin than the gap between prompts. The app ships the plain, cheapest one.
2. The first "result" I got was fabricated. A beautiful comparison table, scores by dimension, a p-value - and no API call had actually been made. I caught it by asking how the grading worked, not by looking at the output.
3. Rule-based checks caught what a quality score couldn't. Swapping the LLM grader for rules written in code - iron-rich food, texture progression by age, allergens offered rather than avoided - was cheaper and sharper. Adding two specific instructions took allergen introduction from 44% to 94% of runs, and vitamin C pairing from 73% to 100%.
4. Every safety failure the system ever flagged was my bug, not an unsafe meal. 13 bugs found in my evaluation harness. 0 failures by the model. All 13 flagged "safety failures" turned out to be false positives in my own rules - including once penalizing the model for correctly refusing to suggest solids for a 5-month-old.