On-device AI & native applications
Turn natural-language requests into app actions.
Connected conversational requests with alarm operations and native iOS scheduling.
Independent product designed and developed by Mathyu Cardozo.
Project
Calarm
Industry
Productivity / consumer applications
Development
1 month of development
Stack
Swift / SwiftUI, Foundation Models, AlarmKit, EventKit, SwiftData, iCloud
Evidence
From natural language to app actions
Mathyu Cardozo · AI Engineer Lead
In production since May 2026, with frequent improvements and new functionality.
The challenge
A conversational request still needed to become a correctly scheduled alarm.
What we built
On-device language interpretation connected application state to native iOS scheduling.
What changed
A published iOS app with documented maintenance; no adoption or revenue result supplied.
Context and problem
Natural-language alarm requests needed to become defined app actions connected to native scheduling.
What we built
Implementation at a glance
Interpret a defined natural-language request.
Connect interpretation with alarm operations and native scheduling.
Maintain the published product after release.
Implementation details
A conversational response alone does not schedule an alarm. The request must become a defined application change, use the right date and recurrence, and reach the system responsible for the alert. Calendar context, local records and shared changes introduce additional responsibilities that must remain clear.
Interpret a defined request
The declared assistant uses on-device Foundation Models to interpret alarm requests. Spoken or typed requests are supported in Spanish and English.
Connect interpretation to execution
The app manages alarm operations, EventKit supplies calendar context and AlarmKit handles native scheduling. SwiftData and iCloud support the declared persistence and synchronization responsibilities.
Maintain the product after release
The supplied May–October 2026 history records 14 versions, including fixes for recurrence, time interpretation and propagation of assistant changes to shared alarms.
Results and evidence
The evidence establishes a published iOS product and specific maintenance work. Local model inference avoids a hosted per-query model fee for that interpretation step; development, distribution and synchronization remain separate costs. Version count does not establish reliability or adoption, and the source provides no download, retention or revenue outcome.
The assistant requires a compatible Apple Intelligence device. iCloud synchronization has separate network boundaries.
Relevance and PDF contents
Where this pattern applies
This pattern is relevant when an assistant should carry out a defined action inside an application. Start with the permitted action, the record it changes, the system that executes it and a clear confirmation. The PDF explains those boundaries and the product changes visible in the supplied release history.
Inside the full PDF
See how language interpretation, application state and native scheduling connect inside Calarm.
Model, app and operating-system responsibilities
Calendar, local records and synchronization boundaries
Examples from the product release history
AI-action readiness worksheet
Free PDF · 8 pages · Includes a readiness worksheet.
On this page
Explore how language interpretation, application state and native scheduling connect. Use the worksheet to assess a similar AI-to-action workflow.
Next case