Built an AI dog translator. It uses a small convolutional neural network to detect a dog in the camera frame, acting as a trigger so heavy AI models only run when needed. Once a dog is present, the app queries a vision language model on the phone and asks questions about posture, nearby objects, body language, other animals, emotions, actions, and overall mood. The microphone listens in parallel, and a bark detector triggers audio slicing into small segments. A bark analysis model, trained on thousands of dog barks, predicts the mood and tone of each bark. Finally, a large on device LLM receives all the vision and audio results, combines them, and generates a realistic guess of the dog’s current thought. The app displays the thought as text and plays audio using a text to speech model.