WhoBuilt

Build Smarter Apps: Tool Calling & AI Orchestration Explained

Mar 12, 2026· 7 min· 380 views·Watch on YouTube

Speaker

Mike, CTO and founder at Callstack, explains tool calling and structured outputs for on-device AI in React Native AI Unpacked.

Tool calling connects models to app capabilities

Tool calling lets a model use functions and APIs to fetch data or take actions, such as checking the weather, querying a database, or sending a notification. Mike says the model must support tool execution and notes that React Native AI Apple supports it.

A calendar demo shows tool use in practice

Mike demonstrates a model calling getCurrentTime and checking a calendar for the day's agenda. He says the calendar result matched the device app, but cautions that returned UTC times need to be localized for the user. With available tools, a model could also assess whether someone will arrive on time for an event.

Structured outputs make responses usable by apps

Structured outputs constrain a model's response to a defined shape, such as a JSON schema with typed fields. Mike demonstrates an object with a person's name, age, and active status, an enum for a color, a Boolean, and arrays. He says these formats can support forms, information extraction, UI generation, and other application logic.

Model behavior depends on the model and settings

Mike shows code that generates a fruit array and says results depend on the model, its parameters, and its architecture. In the demo, repeated runs returned different lists, though the Apple device model favored apples.

Takeaways

  • Check that the chosen on-device model supports tool execution before designing workflows around tool calls.
  • Localize timestamps returned by tools before showing them to users.
  • Use structured outputs when app logic needs predictable, typed data.
  • Test model outputs across runs and settings; sampling parameters do not guarantee a particular result.

Mentioned in this talk

Search inside this talk

What was claimed 5

Related talks