Sentinel AI

Sentinel is a personnel wellbeing monitoring prototype that brings together a live facial-expression analysis tool and an operational dashboard for reviewing wellbeing indicators, prioritizing follow-up, and tracking welfare actions. Its goal is to help teams spot potential stress signals and organize supportive responses; it is not a diagnostic or clinical system. The monitoring view uses a webcam feed and a Python backend built with Flask, OpenCV, and TensorFlow. The backend detects faces and runs a local facial-expression model to estimate one of seven categories: angry, disgust, fear, happy, sad, surprise, or neutral. It can also connect to Gemini Vision when configured with an API key. The interface presents the results as emotion and stress indicators, with a legend that distinguishes baseline and potential stress markers. The React dashboard provides several views for working with this information. The overview presents personnel and risk summaries, stress trends, unit comparisons, intervention charts, priority alerts, and a roster preview. The personnel section supports reviewing and registering personnel records. Dedicated pages provide a larger live-monitoring view, alert management, analytics, welfare recommendations, and settings. Users can acknowledge or resolve alerts, approve pending action plans, inspect individual personnel details, and export information through the dashboard controls. Sentinel is implemented with React, TypeScript, and Vite on the frontend, with a Flask service for model inference. The interface uses charts to make aggregate patterns easier to scan and provides controls for common follow-up workflows. It is designed as a starting point for exploring how computer-vision signals might fit into a broader personnel-welfare process. The project should be treated as a prototype. Much of the dashboard’s roster, alert, recommendation, and aggregate analytics content is initialized from local sample data; the presence of a live inference backend does not mean these views are connected to a complete, persistent operational data system. Facial-expression models can misread expressions and context, and visible emotion is not a reliable measure of a person’s mental state. Any real-world use would need careful validation, privacy protections, informed consent, human review, and safeguards against using estimates as the sole basis for decisions.