I need a deep-learning pipeline that automatically spots human faces in images or live video and then classifies the dominant emotion shown (happy, sad, angry, etc.). The workflow should first localise each face with high accuracy and then run a lightweight yet reliable emotion-recognition module on the cropped face region. You are welcome to choose frameworks you are comfortable with—PyTorch, TensorFlow, Keras, or a well-maintained OpenCV/DNN hybrid—as long as the final model can run on a standard GPU and be exported to ONNX. Pre-trained backbones are fine, but the emotion head must be custom-trained or fine-tuned so the overall system reaches at least 90 % precision and recall on a held-out validation set that we will supply. Deliverables: • Clean, well-commented training code. • Inference script that takes an image or video stream and returns bounding-box coordinates plus an emotion label and confidence score for each detected face. • Saved model weights and ONNX export. • README explaining environment setup, data requirements, and a quick-start demo. • Short report summarising training procedure, hyper-parameters, and evaluation metrics. Acceptance criteria: the demo must run end-to-end on my machine, detect faces under varied lighting, and correctly tag emotions with the agreed-upon accuracy threshold.