Kitchn Pro
Shipping to the App StoreAn iOS meal-planning app I built end-to-end.
React Native · Expo · iOS · LLM tool-calling · routine extraction · vision · evals
Kitchn Pro is an AI-native iOS meal-planning app I built end-to-end. Rather than a chat bolt-on, its assistant takes real actions in the app: classifying user intent and executing in-app changes behind a confirm-and-undo guardrail. That action loop is the product, not a bolted-on chat window.
Tool-calling assistant
Classifies user intent and executes in-app actions behind a confirm-and-undo guardrail, so the model never changes your data without a reversible step.
Model-agnostic eval harness
55 golden test cases run across Anthropic and OpenAI through a provider-abstraction layer, used to route each task to a model based on measured results rather than guesswork.
Conversational routine extraction
Users describe how they usually eat in plain language; the app extracts structured daily routines from the conversation.
Receipt vision
Turns a photo of a receipt into structured, validated data.
How I chose the model
| Metric | Anthropic model | OpenAI model |
|---|---|---|
| Pass rate (/55) | 92.7% | 89.1% |
| Cost per task | $0.00280 | $0.00025 |
I route to Anthropic where accuracy matters most. The cost difference is negligible at current volume, and OpenAI remains a cheaper fallback if cost becomes the priority.
Raw eval output
