Argus

A native macOS / iOS app for browsing Tesla Sentry & dashcam exports. Imports a USB drive's SavedClips/SentryClips folders, plays all four cameras in sync, runs on-device computer-vision analysis to surface noteworthy events, and writes plain-language summaries using Apple's on-device foundation model.

Built with SwiftUI, SwiftData, AVFoundation, Vision, MapKit, and (where available) the FoundationModels framework.

Features

• Multi-cam synchronized playback — Front / Left / Right / Rear in a 2×2 grid driven by a single scrubber. Tap a tile to focus, scrub with ← / → / J / K / L.
• On-device event detection — Vision is used to find humans, vehicles, and license-plate-shaped text on every 5th frame, with monocular distance estimates from the camera's known FOV.
• Behavior tagging — Each event is auto-tagged as Touched / Lingered / Approached / Passing / Vehicle / Noise based on how close someone came and how long they stayed.
• AI summaries — Short natural-language descriptions of each event using Apple Intelligence (FoundationModels) when available, with a deterministic fallback when not.
• Geofenced zones — Drop pins on a map to define Home / Work / etc. Events inside those zones are labeled automatically.
• Map view — All geo-tagged events plotted with color-coded markers per behavior.
• Clip export (macOS) — Save a ±5 s clip around the current scrubber position straight to disk.
• Optional iCloud sync — Events and geofences sync via CloudKit; videos stay local because security-scoped bookmarks aren't portable.

Requirements

• Xcode 16 or newer
• macOS 14+ or iOS 17+
• AI summaries require macOS 26 / iOS 26 with Apple Intelligence enabled. Older OSes fall back to a deterministic summary string.

Getting started

1. Clone the repo and open Argus.xcodeproj in Xcode.
2. Select the Argus scheme and a Mac / iOS destination.
3. Build & run.
4. On first launch, plug in your Tesla USB drive (or mount the export folder), tap Import, and pick the SavedClips or SentryClips directory.
5. Tap Analyze all in the Videos tab to run the Vision pipeline over every imported clip.

Project layout

The codebase is intentionally broken into small files (every file ≤ 200 lines) so any single screen, control, or helper can be tweaked without scrolling through unrelated code.

Argus/
├── Helpers/
│   ├── Analyzer.swift                 # Observable VideoAnalyzer (progress state)
│   ├── DetectionEngine.swift          # Vision pipeline (humans / plates / vehicles)
│   ├── EventSummarizer.swift          # FoundationModels-backed text summaries
│   ├── Geocoder.swift                 # CLGeocoder reverse-geocode
│   ├── TeslaCamera.swift              # Camera ID normalization + FOV table
│   └── TeslaDashcamImporter.swift     # event.json + .mp4 importer
├── Models/
│   ├── Detection.swift                # Detection + DetectionSummary value types
│   ├── Event.swift                    # SwiftData @Model
│   ├── EventTag.swift                 # behavioral tag + classifier
│   ├── Geofence.swift                 # SwiftData @Model
│   └── VideoRecording.swift           # SwiftData @Model + detection markers
└── Views/
    ├── MainView.swift                 # root TabView
    ├── SettingsView.swift             # geofences, bulk ops, iCloud toggle
    ├── Events/
    │   ├── EventsListView.swift       # primary list (search/filter/sort)
    │   ├── EventsListToolbar.swift    # filter menu + import toolbar
    │   ├── EventsImport.swift         # file-importer plumbing
    │   ├── EventRow.swift             # one row in the list
    │   ├── EventChips.swift           # ZoneChip / TagChip / ScoreBadge
    │   ├── EventDetailView.swift      # detail screen composition
    │   ├── EventDetailSections.swift  # summary / metadata / notes sections
    │   └── EventsMapView.swift        # map of geo-tagged events
    ├── Videos/
    │   ├── VideoListView.swift        # grouped clip list + analyze button
    │   ├── VideoRow.swift             # clip row with thumbnail
    │   ├── VideoAnalysisRunner.swift  # batch detection loop
    │   ├── PlayerLayerView.swift      # raw AVPlayerLayer wrapper
    │   ├── PlayerSheet.swift          # single-clip modal player
    │   ├── VideoPlayerView.swift      # AVKit VideoPlayer wrapper
    │   ├── SyncedMultiCamPlayerView.swift              # multi-cam coordinator
    │   ├── SyncedMultiCamPlayerView+Tiles.swift        # grid + tiles UI
    │   ├── SyncedMultiCamPlayerView+Controls.swift     # scrubber + wall clock
    │   ├── SyncedMultiCamPlayerView+Playback.swift     # AVPlayer lifecycle
    │   └── SyncedMultiCamPlayerView+Export.swift       # ±5s clip export
    ├── Settings/
    │   ├── GeofencePickerSheet.swift  # add-zone modal
    │   └── SettingsBulkActions.swift  # recompute zones / dedupe helpers
    └── Style/
        ├── LiquidGlassStyle.swift     # liquidGlassChip / liquidGlassCard
        └── SectionCard.swift          # reusable titled card

Searchable comment tags

Every view file uses short uppercase keyword comments so you can jump straight to whatever you want to tweak. Search the project for any of these:

Keyword	What it marks
UI:	A visible component (row, card, chip, badge, tile, etc.)
LAYOUT:	Padding, spacing, frames, sizes — anything controlling position / dimensions
COLOR:	A color choice — fills, tints, borders, foreground styles
FONT:	Font size, weight, or family
TEXT:	User-visible strings — change these to relabel things
ICON:	SF Symbol choices
BUTTON:	Tap targets and their actions
TUNING:	Adjustable thresholds, magic numbers, detection knobs
PLAYBACK:	AVPlayer lifecycle / seek / play / pause

Examples:

• Want to rename the "Events" tab? Search TEXT: change labels here.
• Want to widen the multi-cam grid? Search LAYOUT: Max width of the 2x2 grid.
• Want a different badge color? Search COLOR: marker tint per tag or COLOR:chip.
• Want to tweak how sensitive the proximity detector is? Search TUNING: lower = more sensitive.

Architecture notes

• SwiftData stores split in two: MetadataStore (Events + Geofences) is CloudKit-syncable, VideosStore is local-only because video URLs depend on security-scoped bookmarks. See ArgusApp.swift.
• Multi-cam sync uses one earliest-start anchor time + per-camera offsets. A periodic time observer on a "primary" camera (preferring Front) drives the shared scrubber position; seek translates the global position back to each camera's local time, clamping or pausing at boundaries.
• Detection runs off-main via Task.detached(priority: .userInitiated), with the VideoAnalyzer @Observable class only holding the progress state on the main actor.
• AI summaries use FoundationModels.SystemLanguageModel.default when .available; otherwise they fall back to a deterministic concatenation of the facts that would have been sent to the model. No data leaves the device.
• License plate filter is intentionally conservative — 4–8 chars, uppercase alphanumeric, must contain both a digit and a letter. False positives from random text in scenes are common otherwise.

Keyboard shortcuts (multi-cam player)

Key	Action
Space / K	Play / Pause
←	Back 5 s
→	Forward 5 s
J	Back 10 s
L	Forward 10 s

Privacy

Everything — Vision detection, license-plate text recognition, and AI summaries — runs on-device. No frames, summaries, or coordinates ever leave the machine. Reverse geocoding goes through Apple's CLGeocoder, which is the only network call the app makes.

License

This is a personal project. No license is currently provided. All rights are reserved unless specified otherwise in the future.