Glean

Glean uses Apple Intelligence to turn handwritten pages into structured notes.

Snap a page. Five parallel Foundation Model passes extract decisions, action items, open questions, and tags. Everything runs on the phone.

No cloud.
No API keys.
No data leaves the device.

Before	After

_{A handwritten page, captured and parsed by Glean on-device.}

Lecture notes, whiteboards, sticky-note walls, the back of a napkin. Glean reads them with on-device Apple Intelligence and saves each capture as a structured note: summary, decisions, action items, open questions, tags. Nothing leaves the device.

Status: in development. App Store submission planned.

---

Demo

glean-flow.mp4

A full capture, from photo to saved note.

---

Highlights

• Apple Intelligence, parallelized. Glean runs five Foundation Model passes concurrently on the device, one per output field. Tuned around the 3B model's accuracy limits on multi-field output.

• A deterministic safety net. Every model output is filtered by word-overlap against the source text and NLEmbedding-based near-duplicate detection before reaching storage. Hallucinations don't ship.

• A measured image pipeline. Thumbnails decode through CGImageSourceCreateThumbnailAtIndex inside autoreleasepool brackets. A 20-note library stays under 30 MB resident.

• Zero network, photo to saved note. OCR, parsing, dedup, storage. All on the phone.

• Native iOS 26. Pure SwiftUI, SwiftData, Liquid Glass, Foundation Models, Vision, NaturalLanguage. No third-party dependencies.

---

The app

Library. Search, sort, pin, swipe to delete. Multi-select edit mode for bulk pin or delete. Pinned notes stay at the top.

Note detail. Every field is editable in place. Action items open a focused editor with date and urgency. Questions can be answered inline.

Source viewer. Pinch-zoom up to 4×. Share or save to Photos.

---

Architecture

flowchart TB
    subgraph UI["SwiftUI views"]
        L[Library]
        C["Capture<br/>VisionKit:<br/>VNDocumentCameraViewController"]
        P[Processing]
        D[NoteDetail]
        S[SourceImage]
    end
    subgraph Data["SwiftData @Model"]
        N[Note]
        AI[ActionItem]
        OQ[OpenQuestion]
    end
    subgraph Pipeline["Pipeline @ModelActor"]
        CP[CaptureProcessor]
        VS[Vision: RecognizeTextRequest]
        FM[CaptureAI: 5 parallel<br/>Foundation Model passes]
        VL[Validation:<br/>word-overlap + NLEmbedding]
        TN[Thumbnail: ImageIO downsample]
    end
    UI -->|"@Query · @Bindable"| Data
    C -->|"imageData"| CP
    CP --> VS --> FM --> VL
    CP --> TN
    VL -->|"insert + save"| Data
    P -->|"observes AsyncStream"| CP

Loading

Views never reach into the pipeline. The capture flow hands imageData to a @ModelActor CaptureProcessor. The actor streams ProcessingEvents back via AsyncStream for ProcessingView to render phase transitions. When it finishes, it inserts the assembled Note plus child ActionItems and OpenQuestions in a single autoreleasepool-wrapped commit.

The pipeline

flowchart LR
    A[Captured photo] --> B[Vision OCR<br/>EXIF orientation]
    B --> C{Pre-parse<br/>section + @-name routing}
    C --> D[Title + Summary]
    C --> E[Decisions]
    C --> F[Action items<br/>+ DateLookupTool]
    C --> G[Open questions]
    C --> T[Tags]
    D & E & F & G & T --> H[Validator<br/>word-overlap ≥ 0.55]
    H --> I[Dedup<br/>NLEmbedding cosine ≥ 0.85]
    I --> J[(SwiftData)]

Loading

A deterministic pre-parse routes each generation to the relevant section of the OCR (text under DECIDED: for decisions, TODO: for actions, and so on) and attaches an @-name constraint to the actions pass when explicit owners are present. On devices without Apple Intelligence, the pipeline falls back to plain OCR and surfaces a "Plain OCR" badge on the note.

---

Tech stack

SwiftUI · SwiftData · Foundation Models · Vision · VisionKit · NaturalLanguage · ImageIO · iOS 26 · Liquid Glass

No third-party dependencies.

---

Privacy

• Notes never leave the device. No server, no cloud LLM fallback, no analytics, no crash reporting to third parties.
• No accounts. No identifier tied to you.
• Storage is local: SwiftData (SQLite plus external-storage blobs) in the app's Application Support directory.

---

Roadmap

• Tasks view. Aggregate action items across all notes into a sorted, filterable list (by owner, due date, urgency).
• Multi-page notes. Capture and store a multi-page scan as a single note instead of taking only the first page.
• iPad and macOS. Multi-column layouts. macOS via Designed-for-iPad initially, native Catalyst later.
• iCloud sync via CloudKit. End-to-end encrypted, Apple-ID-based, zero third-party identity. Preserves the on-device-first promise.
• Share extension and Shortcuts. Capture from any app, automate via Siri.

---

Requirements

• iPhone with Apple Intelligence support for full structured extraction. Older devices fall back to plain-OCR mode.
• iOS 26.0+
• Xcode 16.0+ to build.

---

License

All rights reserved. See LICENSE.