kaopanboonyuen/saie2026

SAIE2026 is the official workshop repository for a course titled AI in the Real World: Trade-offs Behind Fast and Scalable Object Detection. It is part of Super AI Engineer Thailand 2026, a national training program for the next wave of AI engineers. The author is Teerapong Panboonyuen from Chulalongkorn University in Thailand, and the materials are aimed at students who want to take object detection beyond a textbook example and into something that could actually run in a real product. Object detection is the task of finding things inside images or video frames and putting boxes around them. Most tutorials only care about getting the highest accuracy score. This workshop is different because it focuses on the engineering trade-offs that show up once you try to deploy a detector on a real device. Topics include latency versus accuracy, model size versus performance, the limits of running on small edge hardware, and how to handle many camera streams at once. The repository is organised into three main folders. The slides folder holds the lecture PDF. The notebooks folder holds two Google Colab notebooks, one student version and one with solutions filled in. There is also a datasets folder for the working data. The README links out to the Colab notebooks directly, so students can open them in the browser without setting up a local Python environment. The workshop is split into six labs. Lab 1 is baseline profiling, where students measure speed, FLOPs, and accuracy of YOLO variants. Lab 2 covers structured pruning, which means cutting parts of the model that contribute little. Lab 3 covers quantization, which means running the model in lower numerical precision such as FP16 or INT8 to make it faster. Lab 4 covers knowledge distillation, where a small student model learns from a larger teacher model. Lab 5 covers multi-scale detection heads and backbone design. Lab 6 covers multi-camera scalability and deployment, ending with a hackathon challenge. The README also lists a useful set of external links for NVIDIA Jetson optimization, TensorRT, and DeepStream, which are the tools commonly used when running detectors on edge hardware. The license is marked as Educational. The repo is mostly Jupyter Notebook by language tag, which fits its role as teaching material.