I’m developing a Flutter application that must run completely on the user’s device. Using TensorFlow Lite together with MediaPipe, the app should: • accept images taken directly from the camera or selected from the gallery • perform all processing offline, without any server calls • classify the image, return a set of labels, and generate a short auto-caption in real time I will supply UI mock-ups; what I need from you is the full integration of a suitable TFLite model (or a pair of models, if one is better for captioning) and the MediaPipe image pipeline, plus clean Dart code that exposes a simple method such as classifyImage(File img). Final output should include the Flutter project, the model files, brief setup notes, and a README that explains how to replace or retrain the model later. Acceptance criteria – App builds and runs on Android and iOS simulators and at least one physical device. – Processing stays entirely on-device; airplane-mode testing must still return labels and captions. – Average inference time on a mid-range phone ≤ 300 ms for 224×224 inputs. – Caption is returned in natural language, not just a label list. If you have previous work with TFLite, MediaPipe, or similar on-device ML solutions, please share a short example or repo link so I can gauge fit and jump straight into the code review phase.