Meal Decider
A 'what should I eat?' app where most decisions use no LLM at all, and it measures how many.
idea
Summary
A planned few-tap interview that filters real nearby restaurants against preferences such as no chains, budget, and walk time, and returns a top pick with a pickup or reservation link. Typed decisions go to code; an LLM parses free text. Every model call is planned to go through my Model Router so cost per recommendation can be logged.
Problem
I lose time every day deciding what to eat, and generic recommendations often miss my preferences.
AI Usage
Planned: Claude Haiku 4.5 via Model Router turns free-text answers into typed constraints. Deliberately not AI: the interview, place search (Places API), chain detection, filters, ranking, and links. The app never books or pays; a human confirms.
- Plan built from two real decisions made by hand, with every step classified as code or LLM.
- Headline metric planned: share of recommendations that needed zero model calls.
Stack
Python, Anthropic SDK, Model Router, Google Places API