This repository contains a collection of real, production-grade demo implementations showcasing Apple’s iOS 26 Foundation Models (on-device LLM). It is designed as a comprehensive playground for experimenting with:

• On-device language models
• Generable (structured generation)
• Tool (agent-style function calling)
• Transcript (AI workflow timeline)
• HealthKit & EventKit real-world data integration
• Fax / Scan / PDF intelligent automation
• Dynamic schemas & configurable workflows

This project demonstrates how to build real products with Apple’s on-device AI—not just chatbots.

Demonstrates Apple's on-device LLM capabilities introduced in iOS 26:

• Low latency (<100ms local inference)
• Fully private (no data leaves the device)
• Native integration with app-level workflows
• Built-in guardrails for safety

Includes demos for:

• Prompting
• Instruction design
• Safety boundaries
• Streaming generation
• Chat session orchestration

Generable lets developers build strong schema-driven models:

• Define Swift structs that the model must fill
• Replace “LLM hallucinated JSON” with strict, type-safe results
• Perfect for UI forms, workflows, automation logic

Included demos:

• GenerableWorkflow Demo
• DynamicGenerationSchemaDemoView
• DynamicFaxPlanSelection Demo
• DynamicOCRLanguageSelection Demo
• DynamicPDFCompression Demo
• DynamicContactSelection Demo
• DynamicScanPreset Demo
• DynamicWorkflowSchema Demo

This project demonstrates how to turn Swift services into callable LLM tools.

Examples:

• SearchContactTool
• SendFaxTool
• AddCoverPageTool
• SearchFaxTool
• SaveFaxDraftTool
• CropFaxTool

• CorrelateWithRealWorldDataTool → Calls real HealthKit + EventKit

  • Reads sleep data
  • Reads heart rate
  • Reads calendar events
  • Correlates real-world stressors with user dream emotions

All tool calls are automatically shown in the Transcript UI.

Transcript is one of the most important parts of iOS 26’s Foundation Models.

We provide two production-ready implementations:

ChatGPT-style UI with:

• User bubbles
• AI response bubbles
• Tool call bubbles (with tool badges)
• Tool output bubbles
• Streaming responses
• Auto-scroll
• Message grouping

Developer-facing console to debug:

• Tool sequence
• Argument correctness
• Output inspection
• Model decision steps

Transcript entries include:

• .instructions
• .prompt
• .toolCalls
• .toolOutput
• .response

We include a real production-ready version of HealthKit + EventKit integration:

• Requests authorization
• Reads sleep analysis
• Reads heart rate
• Computes pressure nights
• Fully async/await

• Reads real calendar events for “last week”
• Extracts meaningful stress-related events

Used in the dream-context emotional analysis Demo.

A complete end-to-end example:

• User asks: “Why were my dreams so anxious last week?”

• AI is required (via Instructions) to call: correlateWithRealWorldData

• Tool fetches HealthKit + Calendar real data

• Returns structured RealWorldContext

• Model generates empathetic wellness insights based on actual user life events

• Perfect example of combining:

  • LLM reasoning
  • Real Sensor Data
  • Structured output
  • Tool chaining

UI file:

AIHubEngine acts as the central orchestrator:

• Manages session lifecycle
• Registers all tools
• Handles streaming
• Publishes Transcript entries
• Exposes UI-friendly state
• Builds multi-module AI workflows (Fax, Contacts, Dream, PDF, Scan)

This is the “brain” of the entire system.

Root
 ├── AIHub/
 │    ├── AIHubEngine.swift
 │    ├── Tool/
 │    ├── Services/
 │    ├── TranscriptUI/
 │    └── ChatUI/
 │
 ├── Demos/
 │    ├── FoundationModel Demo
 │    ├── Prompt Demo
 │    ├── Safety Demo
 │    ├── ToolFaxScan Demo
 │    ├── WorkflowToolTranscript Demo
 │    ├── Dynamic Workflow Demos (Generable)
 │    └── DreamContextAnalyzerView.swift
 │
 ├── HealthKit/
 │    ├── HealthKitManager.swift
 │    ├── CalendarManager.swift
 │    ├── RealWorldContext.swift
 │
 ├── FaxServices/
 │    ├── FaxService.swift
 │    ├── ContactsService.swift
 │    ├── CoverPageService.swift
 │    └── Tools/*.swift
 │
 └── README.md  ← (this file)

• Xcode 16+

• iOS 26+ device (HealthKit not supported in Simulator)

• Enable in Signing & Capabilities:

  • HealthKit
  • iCloud (optional)
  • Contacts (if you plug deeper)
  • Calendars

Ensure the following keys exist in Info.plist:

<key>NSHealthShareUsageDescription</key>
<string>App uses your health data to build local AI insights.</string>

<key>NSHealthUpdateUsageDescription</key>
<string>App uses health data to improve dream & wellness insights.</string>

<key>NSCalendarsUsageDescription</key>
<string>App reads calendar events to correlate real-world stress factors.</string>

1. Install to a real iPhone (Simulator won't support HealthKit)

2. Launch the app

3. Navigate through the Demos from the main menu

4. Try:

   • Generating structured output
   • Calling tools
   • Observing Transcript
   • Running Dream Context Analyzer
   • Fax AI workflow demo

User → “Help me send this fax to my accountant”
Model → Plans workflow:
         1) searchContact
         2) addCoverPage
         3) sendFax
         4) saveFaxDraft (fallback)

DynamicFaxPlanSelection
 → Changes fields based on country/user context
 → Schema passed to model
 → Model outputs strict Generable struct

Tool: correlateWithRealWorldData(period: "last_week")
 → Reads sleep analysis
 → Reads heart rate
 → Reads calendar events
 → Returns RealWorldContext summary
 → Model produces emotional explanation

• PDF semantic segmentation (on-device)
• On-device OCR auto-language selection
• Tool router for multi-domain AI orchestration
• Local RAG (device-only embeddings)
• AI-powered scanning skills (cropping, enhancement)
• Fax analytics + workflow optimization
• Jet AI Hub: unified AI control center for all Jet apps

This project is for research, demonstration, and education. Commercial usage should ensure compliance with Apple policies, HIPAA/PII rules, and local regulations.