sud747/vizualai

Purpose and deployment shape

VizualAI recognizes foods in tray images and returns an itemized bill. It is a multi-client repository centered on a Python FastAPI inference service, with packaged ML weights and Flutter/native Android clients. The API is intended to run locally on port 8001; Docker support is present via Dockerfile.

Backend API entry points

The documented service entry point is backend/food_app.py, which exposes REST endpoints and hosts model inference. backend/app.py is a second backend application/module and should be assessed before consolidating server startup behavior. A static browser-facing dashboard is co-located with the backend at backend/static/food_ui.html and backend/static/index.html.

Vision inference pipeline

The core recognition boundary is the two-stage pipeline in backend/yolo_pipeline.py:

1. YOLOv8 detects candidate tray objects and bounding boxes.

2. Regions are cropped/preprocessed and classified by fine-tuned MobileNetV2.

3. Custom post-processing/NMS removes duplicates.

4. Recognized classes feed pricing/billing returned by the API.

The deployable model artifacts are models/yolov8n.pt and models/best_food_classifier.pth. backend/detect_food.py provides a likely CLI-oriented invocation path separate from the HTTP service.

Model lifecycle

Inference consumes checked-in model weights from models/. Training is separated from serving in backend/train_classifier.py, creating a material lifecycle boundary: training or replacing classifier weights must preserve the class-label and pricing assumptions used by the runtime pipeline. GPU acceleration is optional through CUDA-compatible PyTorch; CPU execution remains supported.

Flutter client flow

The principal cross-platform client starts at apps/flutter/lib/main.dart and targets Android, iOS, web, Windows, macOS, and Linux through Flutter platform hosts. UI flows are organized around start, scanning, and billing screens: apps/flutter/lib/screens/start_screen.dart, apps/flutter/lib/screens/scanner_screen.dart, and apps/flutter/lib/screens/billing_screen.dart. Scanner and payment state/control flows use BLoC boundaries in apps/flutter/lib/blocs/scanner/scanner_bloc.dart and apps/flutter/lib/blocs/payment/payment_bloc.dart.

Client-to-service and billing boundary

The Flutter repository layer, apps/flutter/lib/repositories/food_repository.dart, is the client-side boundary for food/API data. API endpoint configuration is isolated in apps/flutter/lib/config/api_config.dart. Domain payloads are represented by apps/flutter/lib/models/food_item.dart and apps/flutter/lib/models/bill.dart; the expected flow is scan image → API prediction → recognized items → bill display/payment state.

Native Android client

A separate Kotlin/Jetpack Compose Android implementation exists, entered through apps/android/app/src/main/java/com/example/foodsight/MainActivity.kt. It uses navigation via Navigation.kt, presents its primary UI in ui/main/MainScreen.kt, and separates screen state in ui/main/MainScreenViewModel.kt. Its data boundary is data/DataRepository.kt; it should be treated as an independent client implementation rather than a Flutter platform wrapper.

Infrastructure and external dependencies

Runtime dependencies include FastAPI/Uvicorn, Ultralytics YOLOv8, PyTorch/Torchvision, Pillow, and multipart upload handling, declared at repository level in requirements.txt. External execution dependencies are Python 3.8+, optional NVIDIA CUDA/GPU support, Flutter SDK for the cross-platform app, and Android Gradle tooling for the native client.