A repair shop in Abu Dhabi takes forty calls a day and writes down none of them. The customer's name, the device, the day they need it back — all of it lives in the owner's head until it doesn't, and the lead is gone.
Sayd is a five-second voice note that becomes a customer record and a scheduled follow-up. The owner speaks — in Arabic, English, or the mix of both people actually use — and the app transcribes it, pulls out the name, the request, the phone number and when to chase, and puts a reminder on the calendar. Every step runs on the phone. There is no server, no API key, and no account.
Capture, the Today list sorted by what is about to be lost, a lead's detail, and the whole app in Arabic with a right-to-left layout.
Piloted with BILAL GSM, a mobile repair shop in Abu Dhabi. The owner's day is his hands: he is holding a phone open with a spudger when the next customer calls. Typing a name into a CRM is not something that happens, and by evening he can remember maybe half of who wanted what.
Three constraints fall out of that, and they shape the whole app:
- Capture has to cost nothing. One tap, five seconds of speech, no typing. If logging a lead is slower than not logging it, nothing gets logged.
- The list has to be sorted by risk, not by time. "What did I write down first" is useless. "Who am I about to lose" is the only question worth asking.
- The customer list cannot leave the phone. It is the shop's commercial asset. A privacy promise that quietly degrades is not a promise.
SFSpeechRecognizer runs with requiresOnDeviceRecognition = true, which makes it
refuse when on-device recognition is unavailable rather than silently posting the
audio to Apple's servers. The transcript then goes to Apple's on-device
SystemLanguageModel through a @Generable struct, so the model is constrained to
return the four fields the app needs and cannot hand back malformed JSON. When
Apple Intelligence is unavailable — an older iPhone, the model still downloading —
LeadExtractor.heuristic takes over with plain string rules. Capture never fails.
Nothing is written to the database until the owner has looked at it. An extraction that is right nine times in ten still loses one customer in ten; a one-tap confirmation is cheaper than a lost lead.
The best version of this app is one the owner rarely opens. The medium widget puts the leads going cold — and the week's follow-up rate — on the home screen, so glancing is enough.
The widget, the Share Extension and the Siri intent all read the same SwiftData store through an App Group, so they cannot drift apart.
brew install xcodegen
xcodegen generate # produces Sayd.xcodeproj from project.yml
open Sayd.xcodeprojThen set your team ID in project.yml (DEVELOPMENT_TEAM) and confirm the App
Group group.com.sayd.shared exists on all three targets.
- Xcode 26 or later
- iOS 26 deployment target
- A physical iPhone — the Speech framework's on-device recogniser and the Foundation Models framework are unreliable or unavailable in the Simulator
- Apple Intelligence enabled, for extraction (there is a rule-based fallback if not)
- Xcode → New Project → iOS → App
- Product Name:
Sayd - Interface: SwiftUI, Language: Swift, Storage: SwiftData
- Product Name:
- Delete the generated
ContentView.swiftandSaydApp.swift. - Drag the
Sayd/folder from this repo into the project navigator (check Copy items if needed, Create groups). - File → New Target → Widget Extension, name it
SaydWidget, uncheck "Include Configuration Intent". Replace its generated Swift file withSaydWidget/SaydWidget.swift. - File → New Target → Share Extension, name it
SaydShareExtension, and replace its generated files with the ones inSaydShareExtension/. - Add
Sayd/Resources/Localizable.xcstringsto the app target.
On the Sayd target → Signing & Capabilities:
- + Capability → App Groups →
group.com.sayd.shared(must matchSaydStore.appGroupID, and must be added to all three targets — without it the widget silently opens its own empty database) - + Capability → Background Modes → Remote notifications (optional)
Add the same App Group to the SaydWidget and SaydShareExtension targets.
In Info.plist add:
| Key | Value |
|---|---|
NSMicrophoneUsageDescription |
Sayd records a short voice note so you can log a customer without typing. |
NSSpeechRecognitionUsageDescription |
Speech is transcribed entirely on your device to turn your note into a lead. |
Project → Info → Localizations → add Arabic and Turkish. To test the RTL layout: Scheme → Run → Options → App Language → Arabic, or use Right-to-Left Pseudolanguage.
Sayd/
├─ Models/
│ └─ Lead.swift SwiftData model; owns the "going cold" rule
├─ Services/
│ ├─ SpeechCaptureService.swift AVAudioEngine + Speech, on-device forced
│ ├─ LeadExtractor.swift Foundation Models @Generable + rule-based fallback
│ ├─ SaydStore.swift shared App Group container (app + widget + share)
│ └─ FollowUpScheduler.swift UNCalendarNotificationTrigger per lead
├─ Views/
│ ├─ CaptureView.swift mic, live waveform, language toggle
│ ├─ TodayView.swift going-cold vs scheduled, week stats
│ ├─ LeadDetailView.swift extracted fields, call / WhatsApp, reschedule
│ └─ ConfirmLeadView.swift human check before anything is saved
├─ Intents/
│ └─ AddLeadIntent.swift "Hey Siri, add a lead"
└─ Resources/
└─ Localizable.xcstrings Arabic, English, Turkish
SaydWidget/
├─ SaydWidget.swift medium widget: leads going cold + follow-up rate
└─ Info.plist widgetkit extension point
SaydShareExtension/
├─ ShareViewController.swift forwarded WhatsApp message → lead
└─ Info.plist share extension point + activation rule
SaydTests/
└─ SaydTests.swift Swift Testing: cold-lead rules + extraction fallback
On-device is a constraint, not a feature. requiresOnDeviceRecognition = true
makes the recogniser fail rather than silently fall back to Apple's servers. A
repair shop's customer list is commercially sensitive; a privacy promise that
degrades quietly is not a promise.
The model never writes to the database directly. ConfirmLeadView sits between
extraction and persistence. An extraction model that is right 90% of the time still
loses one customer in ten — a one-tap confirmation is cheaper than a lost lead.
A phone number is only kept if it was actually said. LeadExtractor.sanitise
re-checks any number the model returns against the digits in the transcript and
drops it otherwise. A hallucinated number is worse than a blank field: the owner
dials it, reaches a stranger, and stops trusting the app.
Graceful degradation. LeadExtractor.heuristic runs when Apple Intelligence is
unavailable, so the app still works on an older iPhone. Capture never fails.
"Going cold" is the organising idea. The list is not sorted by creation date but by how close a lead is to being lost. The widget exists so the owner often does not need to open the app at all.
xcodebuild test -scheme Sayd -destination 'platform=iOS,name=<your device>'Swift Testing covers the two things that are genuinely easy to get wrong: the "going cold" boundary conditions, and the extraction fallback — including that a model number is not mistaken for a phone number, and that a number the model invented is thrown away.
- Voice capture, on-device transcription (AR/EN)
- On-device extraction with fallback
- Phone-number extraction with a one-tap call / WhatsApp follow-up
- Follow-up notifications, Today view, lead detail
- Siri App Intent, home-screen widget
- Arabic RTL, English, Turkish localisation
- WhatsApp Share Extension (parse a forwarded message into a lead)
- Unit tests, XcodeGen project spec
- watchOS companion for wrist capture
- Live Activities for follow-up countdowns
- Gulf/Levantine dialect accuracy tuning
MIT — see LICENSE.


