safety-ai

Multi-modal content moderation system. Detects toxic text and unsafe images in real-time using BERT and Computer Vision. Designed for high-performance safety governance.

Python package to accelerate research on generalized out-of-distribution (OOD) detection.

Code for the paper "A Data-Driven Measure of Relative Uncertainty for Misclassification Detection"

Open-source safety and cognitive control infrastructure for physical AI agents — robots, LLM agents, autonomous systems.

A custom YOLOv8 helmet-detection model trained for 100 epochs using a Roboflow dataset. Includes scripts for real-time image and video detection with bounding boxes, confidence scores, and object tracking with persistent IDs for accurate rider safety monitoring.

AI response safety/ethics/accuracy checker - by A0 (currently not maintained)

Scalable, asynchronous multimodal content moderation pipeline with CV, LLM reasoning, and human-in-the-loop review.

AI-powered ambient distress detection prototype for proactive personal safety with future wearable integration.

Safe multi-modal autonomous driving with cross-attention perception and formally verified QP runtime enforcement shield.

Real-time person-tracking with YOLOv8, zone-dwell analytics and automatic alerts.

Real-time fall detection using YOLOv8 pose estimation, geometric posture analysis, timestamped alerts, and PyTorch versus ONNX benchmarking.

A robust, real-time Face Mask Detection system built with Python, OpenCV, and TensorFlow. Features automated environment setup, FPS monitoring, and fail-safe execution.

Real-time computer vision safety monitoring system using OpenCV and PyTorch for industrial hazard detection

Agentic AI system for Nordic wind-turbine manufacturing using AWS Bedrock and Claude, reducing downtime, enforcing safety, and preventing high-cost mechanical failures.

AI-powered safety equipment monitoring system to ensure workplace compliance through real-time detection.

Real-time adaptive learning system for autonomous vehicles using NVIDIA Cosmos Reason2-2B. Generates safety rules from edge cases, achieving 1500x speedup after learning. Built for NVIDIA Cosmos Developer Competition 2026.

YOLOv8-based kitchen utensil detection and classification system for safety-aware assistive environments, with custom dataset and occlusion-aware evaluation.

Code for the Paper "A Functional Data Perspective and Baseline on Multi-Layer Out-of-Distribution Detection"