🔥 NeuroPro 🌿

Real-time AI vision, rendered on the Metal GPU — 100% on-device.

SwiftUI · Metal 3 · Vision · Mossfire Dusk Design System

“AI That Respects Privacy.” Every pixel is processed on the Neural Engine and GPU. Zero cloud, zero telemetry, zero compromise.

---

✨ Highlights

🎯 Object Detection	Animals with bounding boxes + live scene classification chip (indoor · 87%)
🎨 Style Transfer	Real-time cinematic grade via a custom Metal compute kernel
🧍 Segmentation	Person silhouette isolation, rendered as a live mossfire-tinted overlay
📏 Depth Estimation	Honest production status — reports device capability, real depth mapping staged
🎛️ Filters	Sobel Edge Detection & Preprocess (sharpening + contrast) compute shaders
⚡ Performance Mode	Half-resolution render + MetalFX spatial upscaling (A12+ devices)
🌡️ Thermal Governance	Frame-rate governor that watches ProcessInfo.thermalState + camera system pressure
🔒 Privacy First	All inference on-device — no network permission, no analytics, no PII

---

🏗️ Architecture

┌──────────────────────────────────────────────────────────────────────┐
│  AVCaptureSession ── 1080p @ 60 fps ── CVMetalTextureCache (zero-copy) │
│        │                                                              │
│        ├──► PreviewRenderer ──► CAMetalLayer ──► Screen   (lowest latency)│
│        │        ▲   direct draw from capture queue, no main-thread hop │
│        └──► NeuralProcessor (dedicated neural queue)                   │
│                 ├── Vision ── animals · scene · segmentation (10 Hz)   │
│                 └── MetalEngine ── filters · style · texture pool      │
└──────────────────────────────────────────────────────────────────────┘

Why it's fast (verified 2026-08-05, iPhone 15 Pro / iOS 27):

Metric	Value
Capture	1080p 60 FPS (explicit 60fps format selection)
Preview frame time	11.1 ms (was 17.4 ms)
GPU utilization	47 % (was 72 %)
AI pass	5.5 ms @ 10 Hz on a dedicated neural queue
Dropped frames	0

The hot path avoids per-frame allocations and GPU syncs (waitUntilCompleted) via a reusable texture pool; inference never blocks the capture queue.

flowchart LR
    CAM[AVCaptureSession<br/>1080p @ 60fps] --> MTX[CVMetalTextureCache]
    MTX --> RENDER[PreviewRenderer<br/>CAMetalLayer direct]
    RENDER --> SCREEN[Screen]
    MTX --> AI[NeuralProcessor<br/>neural queue · 10Hz]
    AI --> VN[Vision · Neural Engine<br/>animals · scene · segmentation]
    AI --> MTL[MetalEngine · compute<br/>edge · preprocess · style]

Loading

---

📦 Core Modules

File	Responsibility
NeuralVisionProApp.swift	@main — Metal init, AppState, unbuffered console logging
ContentView.swift	Full-bleed camera UI, floating glass chrome, detection overlay
MetalEngine.swift	GPU engine: pipelines, shader dispatch, texture pool, MetalFX scaler
MetalShaders.metal	Compute kernels (edge, preprocess, style) + aspect-fill preview fragment
NeuralProcessor.swift	AI pipeline: Vision requests, 10 Hz throttle, detection hold window
CameraManager.swift	AVFoundation capture, 60fps format selection, thermal/pressure governor
ThermalObserver.swift	Thermal state → safe frame-rate recommendation
MossfireDuskTheme.swift	Mossfire Dusk design system (colors, type, spacing, motion)

---

🚀 Getting Started

Requirements — iPhone with A12 Bionic+, iOS 16+, Xcode 15+ (beta toolchain tested).

# 1. Build for a connected device
xcodebuild -project NeuralVisionPro.xcodeproj -scheme NeuralVisionPro \
  -destination 'id=<DEVICE_UDID>' -derivedDataPath build \
  DEVELOPMENT_TEAM=<TEAM_ID> CODE_SIGN_STYLE=Automatic \
  -allowProvisioningUpdates build

# 2. Install
xcrun devicectl device install app --device <DEVICE_UDID> \
  build/Build/Products/Debug-iphoneos/NeuralVisionPro.app

# 3. Launch
xcrun devicectl device process launch --device <DEVICE_UDID> \
  com.elite.neuralvisionpro --terminate-existing

For a free-team (personal) Apple ID, pass your team ID via DEVELOPMENT_TEAM on the command line — profiles are 7-day and regenerated automatically.

---

🧪 Verification Workflow

Every feature can be driven live and verified on-device:

# Drive the app to a specific state (DEBUG builds only)
xcrun devicectl device process launch --device <UDID> \
  com.elite.neuralvisionpro --terminate-existing \
  -debugMode StyleTransfer -debugFilter EdgeDetection -debugMetrics on

# Stream runtime logs (camera config, frame rate, mode switches)
xcrun devicectl device process launch --device <UDID> --console \
  com.elite.neuralvisionpro

# Pull crash logs after a run
xcrun devicectl device copy from --device <UDID> --source / \
  --destination ./crashlogs --domain-type systemCrashLogs

Launch arg	Values
-debugMode	ObjectDetection, StyleTransfer, Segmentation, DepthEstimation
-debugFilter	None, EdgeDetection, Preprocess
-debugMetrics	on, off

---

🗂️ Project Structure

NeuroPro/
├── NeuralVisionPro/
│   ├── NeuralVisionProApp.swift   # App entry · Metal init · AppState
│   ├── ContentView.swift          # Camera UI · overlays · detection boxes
│   ├── MetalEngine.swift          # GPU engine · pipelines · texture pool
│   ├── MetalShaders.metal         # Compute + preview shaders
│   ├── NeuralProcessor.swift      # AI pipeline · Vision · throttling
│   ├── CameraManager.swift        # Capture · 60fps · thermal/pressure
│   ├── ThermalObserver.swift      # Thermal → frame-rate governor
│   ├── MossfireDuskTheme.swift    # Mossfire Dusk design system
│   ├── Assets.xcassets            # App icon · colors · launch screen
│   └── Info.plist
├── NeuralVisionPro.xcodeproj/
├── *.md                           # Design, build, handoff, research docs
└── project_stats.sh               # Repo stats utility

---

📚 Documentation

Doc	Purpose
ARCHITECTURE.md	Technical deep-dive: shaders, threading, memory
BUILD_GUIDE.md	Step-by-step build & signing guide
DEVELOPER_HANDOFF.md	Phase-by-phase build-out plan + session wrap-up
BRANDING.md	Mossfire Dusk brand & design language
PROJECT_SUMMARY.md	Executive overview
MARKET_RESEARCH.md	Market analysis & competitive landscape
UX_ENHANCEMENTS.md	UX roadmap & design rationale
SECURITY.md	Privacy & security posture
CHANGELOG.md	Release history

---

🗺️ Roadmap

• Phase 0–2 — Clean build, device deployment, crash-free operation, performance (60 fps / 11 ms)
• Detection — Animal boxes, scene classification, confidence thresholds, anti-flicker hold
• UX — Full-bleed feed, floating glass chrome, mode toasts, dismissible HUD
• Onboarding — 3-screen tutorial, gesture navigation, contextual hints
• General object detection — bundled Core ML model (cars, chairs, …) — staged, model intentionally not added to keep the app lean
• Real depth mapping — LiDAR/dual-camera depth stream
• App Store — paid Developer Program, privacy manifest, iconography

---

🤝 Contributing

Proprietary portfolio/investment project. Technical questions via Issues; partnership inquiries through GitHub.

📄 License

Proprietary — © 2025–2026. All rights reserved. Mossfire Dusk color palette is open for community use.

🙏 Acknowledgments

Apple (Metal · Core ML · Vision), the OpenCV/Vision research community, and the warm glow of a moss-forest sunset at dusk.

Built with 🔥 Metal · 🌿 SwiftUI · 🌑 Mossfire Dusk

GitHub · @kakashi3lite

Made with 🔥 and 🌿

Elite iOS • Metal GPU • On-Device AI • Privacy-First