Seoul Vista is a Flutter app showing live subway, bus and Han River bus positions on a 3D map of Seoul. It offers last-train alerts, Live Activity and AI day plans based on social content. It is released on the App Store.

Core experience

• Interpolate between five-minute location updates to present smooth train motion.
• Select trains and stations for camera tracking, arrival details and delay context.
• Extract places and mood from images, text and URLs, then generate three day-plan options.
• Deliver platform-specific iOS and Android visual languages from one Flutter codebase.

Ownership

The work integrates 3D map abstraction, public-data services, simulation and interpolation, AI services, adaptive UI and delivery.

Current status: a working application and web demo are public. Roadmap items are separated by implementation state.

1. Application and platform layers

The App Shell owns routing, theme and dependency lifetime, while Presentation composes maps, panels, search and itinerary views. iOS Glass and Android Material 3 diverge in Adaptive UI. Views do not call service implementations or Mapbox directly. Supabase authentication and storage sit behind adapters so remote sessions remain separate from view state.

2. City-data normalization

Data/Models converts Seoul Open Data subway, weather and closure responses plus OSM GeoJSON rails into common models. Station, line and coordinate contracts join live responses to static geometry. When an API is empty or rate-limited, demo fixtures provide the same service interface and preserve the presentation path.

3. Train position and rendering path

TrainSimulator combines low-frequency positions, schedules and rail segments. TrainInterpolator estimates progress between updates. The result becomes source and layer updates through IMapController, and the Mapbox adapter renders GeoJSON and markers per frame. Separating refresh cadence from render cadence avoids displaying five-minute input as discontinuous jumps.

4. AI place and itinerary path

Gemini Vision structures social images and text into place candidates and evidence. Geocoding turns valid candidates into coordinates and rejects malformed or out-of-Seoul results. A Dijkstra-based planner orders the route and returns markers, polylines and itinerary cards to Presentation.

5. Trust and failure boundaries

Seoul APIs, Gemini, Mapbox and Supabase are independent external boundaries. Adapters normalize loading, empty, error and demo states, while the map contract limits engine-specific changes. AI places and routes remain suggestions and do not replace official safety or real-time operating information.

6. State and data ownership

The app bundle owns static GeoJSON and transit constants, the presentation layer owns search and map viewport state, and Supabase owns rooms, members, votes and shared location. Keeping recent local searches separate leaves readable city information available during a network outage.

7. Performance and failure recovery

Rendering many lines, stations and POIs makes map-source updates and widget rebuilds likely bottlenecks. Viewport-based layer limits, throttled location publishing, per-service timeouts and caches require validation; the last static transit data is the external-API fallback.

8. Security, observability and debt

A debug path that prints the map token in main.dart should be removed. Supabase RLS, owner RPCs and notification functions form trust boundaries, but unified tracing and external-service SLOs are missing. Map FPS and battery cost have not been measured.

Abstract the map engine

Direct Mapbox calls from views would make SDK replacement and testing difficult. A shared IMapController keeps presentation code independent while Mapbox remains the primary implementation. This costs more initial interface work but reduces coupling.

Separate updates from rendering

Public positions arrive every five minutes while the screen requires continuous motion. The app uses schedules and rail geometry for interpolation instead of requesting data each frame. The UI must therefore distinguish estimated motion from observed positions.

Use platform-adaptive UI

iOS translucent glass and Android Material 3 are not collapsed into a lowest-common-denominator theme. Adaptive widgets swap platform chrome and tokens while domain views remain shared.

Separate map lighting from app theme

City lighting follows time of day, while panel theme follows user preference. Independent state allows a daytime city under a dark application theme without conflating the two systems.

Verification targets

• Confirm that line GeoJSON preserves rail order and connectivity.
• Compare Live and Demo modes against the same map and panel contracts.
• Verify that interpolation remains on rail geometry between API updates.
• Confirm functional parity between iOS Glass and Android Material 3 variants.
• Keep API keys outside the repository and inject them through --dart-define and platform settings.

Performance handling

Device profiles adjust FPS and polling intervals, with user overrides available in settings. The repository documents a 60fps target, but sustained frame-rate coverage across a full device matrix is not measured publicly.

Known limits

The client cannot eliminate delayed or missing public data. Interpolation improves visual continuity but is not precise train tracking. Gemini and geocoding outputs remain recommendations that require user review.

Completed

• Mapbox 3D map and map-controller abstraction
• Line, station and train visualization with frame interpolation
• Train and station selection, camera tracking, weather and time lighting
• Email, social and guest authentication with adaptive platform UI
• Gemini Vision social-content analysis and day plans
• Device-aware rendering and polling controls

In progress

• Stronger live-data failure handling and delay detection
• Refinement of saved-place and travel-tab flows

Planned

• Congestion heatmaps and multimodal routing
• Voice assistance and navigation integration
• iOS Live Activities and stronger onboarding
• Permission, privacy and performance validation for store delivery

Out of scope

This is not a safety-critical train-control or dispatch system. It remains an exploration and visualization product that exposes public-data delays and AI uncertainty.