Drunk driving has remained one of the leading causes of road fatalities worldwide, with an estimated 273,000 deaths annually. Current solutions, such as in-vehicle technologies, outreach activities, and traffic cameras, are either invasive, expensive, or insufficient. In response to this urgent threat, I created RoadSense - an AI-powered camera system which uses vehicle behaviour to detect drunk and impaired drivers. It is powered by a Jetson Orin Nano and EMEET 1080p camera, and it includes a 3d printed casing to protect it from environmental circumstances. This system uses 4 behavioural markers to detect impaired drivers following object identification: lane deviation, speed fluctuation, trajectory changes, and proximity violations. Each detection mechanism has been designed using different machine learning models, perspective transformation, and Hough transformation, which allowed the system to properly analyze multiple vehicles at once on a road. Testing was conducted in both the real world and simulations in Grand Theft Auto V. This technology has achieved high accuracy in all of its detection mechanisms. In object detection, the model can detect cars, motorcycles, trucks, and buses with an overall accuracy of 73%. In lane deviation, during sunny conditions, the model can predict 95% of lanes. In trajectory fluctuations, the overall F1 score is in predicting a drunk driver correctly is 97%. In proximity violations, the model can predict tailgating correctly 95% of the time. At a final material cost of $277, RoadSense offers an impactful solution to monitor road safety and reduce human error in enforcement.