Recherche un développeur compétant a la fois dans le développement et dans ia

Détails de l'offre

📋 Final Technical Brief AI-Powered Video Extraction Pipeline, Frame Processing, Auto-Checkout, and Display on a Private Mini-Site Context: Public Feed Automation and Monitoring Script (Snapchat) Project Objective Design a high-performance automated script capable of monitoring the public Snapchat stories of influencer Benoît Chevalier during a media drop at a specific time (12:00 PM Paris time), extracting visual information from a shared credit card using AI in two steps (videos), executing an automated payment to a pre-configured Eneba account, and instantly displaying the codes of purchased Amazon cards on a dedicated mini-site at the end of the process. ⏱️ 1. Time Synchronization and Warm-up • Time Orchestration: The script must be synchronized via the NTP protocol to the Europe/Paris time zone to align with the noon drop time to the millisecond. • Cloud Server Initialization: Power on a GPU-optimized instance (e.g., NVIDIA T4 / A10G) at 11:55 AM. The script must immediately load the Deep Learning models into the GPU's VRAM and perform an initial cold start inference test to eliminate any startup latency. • Automatic Shutdown: Schedule an automatic kill switch at 12:05 PM to shut down the instance and stop billing. 🎥 2. Video Stream Processing by Frames (Video Ingestion) • Capture in RAM: Intercept the video stream (H.264) of the target influencer's Snapchat stories as soon as they are uploaded to the social network's CDN. Decoding must be done directly in RAM (via OpenCV / cv2.VideoCapture), without intermediate writing to disk to maximize speed. • Sampling (Skip Frames): To optimize processing time, apply a sampling rate so that only one image is extracted and sent to the AI ​​every 6 frames (approximately 5 frames per second of video, which is sufficient to capture the map). 🧠 3. AI-Powered Visual Segmentation and Recognition (Computer Vision Layer) • Spatial Localization (ROI): Use a lightweight object detection model (such as YOLOv8-nano or the EasyOCR detector) to locate the bounding box of the credit card on the screen and automatically crop the frame to it to eliminate background noise. • Reading and Extraction: Pass the cropped area through an OCR (Deep Learning) neural network to decode