Executive Summary
to production-ready app
AI-orchestrated
of production code
full UI localization
PDF, PNG, JPEG, GIF, TIFF
iOS + Android
The Project
A mobile-first application for signing and stamping PDF documents on iOS and Android. Users place signatures, stamps, and images directly onto PDF pages, resize and rotate them with touch gestures, and save or share the result — all without uploading anything to a server.
The app also accepts image files (PNG, JPEG, GIF, TIFF) and converts them to A4 PDFs automatically, applying EXIF orientation correction so photos from any camera appear correctly without manual rotation.
No backend. No user accounts. No cloud storage. Everything runs on the device.
The Method
One engineer, 10 days. The same AI orchestration methodology proven on enterprise backend systems — applied to a consumer mobile product.
Seven AI agents covered the full scope: architecture design, Flutter/Dart implementation, platform-native code (Swift for iOS, Kotlin for Android), localization workflow, and documentation. The engineer made all design decisions, reviewed every output, and owned the final product end-to-end.
10 days to a production-ready app with 60+ languages, on-device AI, and support for five file formats across two platforms. Not a prototype — a fully functional product ready for App Store and Google Play submission.
What Was Built
PDF Viewer and Editor
A high-performance PDF viewer with pinch-to-zoom, smooth scrolling, and virtualized page rendering — only visible pages are rendered at any time, keeping memory stable on large documents. Users drag images from a library onto any page; placed images support free-form resize and rotate via touch handles.
Image Library
A persistent image library stored locally with Isar database. Images are validated on import, normalized for EXIF rotation, and available across sessions. Supports adding images from camera roll or file system.
On-Device AI Background Removal
When a user adds a signature or stamp photograph, the app detects whether the image has a uniform background and offers to remove it. The pipeline runs entirely on-device — no network request is made.
The four-stage pipeline: illumination normalization corrects uneven camera lighting using native image processing (Accelerate on iOS, optimized sliding-window filters on Android); perimeter sampling detects uniform backgrounds before invoking ML; ML segmentation uses iOS Vision Framework (VNGenerateForegroundInstanceMaskRequest) on iOS 17+ and ML Kit Subject Segmentation on Android, with binary threshold at 50% confidence for crisp document-quality edges; post-ML cleanup removes paper trapped inside stamp loops using histogram-based dominant color detection and triple criteria (RGB distance, HSV saturation, HSL lightness) to protect ink of any color.
Image-to-PDF Conversion
Opens one or multiple images and converts them to an A4 PDF with correct margins and EXIF normalization. Single-image and multi-image flows supported.
Localization — 60+ Languages
Full UI localization via Flutter's intl package with ARB files for 60+ languages including RTL support for Arabic, Hebrew, and Persian. Translations generated and reviewed with AI — the same orchestration methodology applied to content as to code.
Session Architecture
Each opened document creates an isolated viewer session with independent state: placed images, selection, dirty flag, file source. Riverpod family providers are fully invalidated on session close — opening a new document never bleeds state from a previous one.
Tech Stack
Comparison with Traditional Dev
| Metric | Traditional Team | This Project (AI-Driven) |
|---|---|---|
| Timeline | 2–3 months | 10 days |
| Team size | 3–5 engineers | 1 engineer |
| Platforms | Often iOS-first, Android later | Both simultaneously |
| Localization | Separate sprint | Included — 60+ languages |
| AI features | Separate ML specialist | Included, same engineer |
| Backend | Usually required | Zero — fully on-device |
The AI-driven approach doesn't replace engineering expertise — it multiplies it. Speed and quality depend directly on the orchestrator's qualifications: architecture understanding, problem decomposition, precise prompting, and critical evaluation of AI output. The human stays in control.
Conclusions
A cross-platform mobile app with on-device AI, 60+ languages, and five file format support — delivered in 10 days by one engineer.
The absence of a backend is not a limitation but a design decision: user documents never leave the device, there is nothing to authenticate, nothing to maintain, and no infrastructure cost. The app works offline, always.
The AI orchestration methodology that produced 13 enterprise microservices in 12 days works equally well for a consumer mobile product. Different domain, different stack, same multiplier.