Paradox

## Project Description Multi-Stream Video Intelligence is an AI-powered video surveillance and analytics platform designed to transform traditional CCTV monitoring into an intelligent, automated, and searchable video intelligence system. The platform processes multiple video sources, analyzes visual content, detects and tracks objects, monitors predefined zones, identifies important events, and retrieves relevant footage through natural-language queries. The system integrates four specialized AI models to provide comprehensive video intelligence: - **SAM 3.1:** Object detection, segmentation, and video tracking. - **SigLIP2:** Image embeddings and text-image similarity for visual search. - **BGE-M3:** Dense and sparse text embeddings for semantic and keyword-based retrieval. - **Muse Spark 1.3:** Multimodal intelligence, natural-language understanding, contextual reasoning, and question answering over retrieved video evidence. The three visual and embedding models operate as independent AI inference services on Baseten, each with a dedicated inference endpoint. Muse Spark 1.3 provides the higher-level intelligence layer, interpreting user queries and generating contextual responses using retrieved evidence. The backend coordinates video preprocessing, AI inference, object tracking, event detection, semantic indexing, and evidence generation. Qdrant manages vector retrieval, while structured databases maintain event records, timestamps, camera information, and tracking history. The platform provides a unified dashboard for multi-camera monitoring, intelligent video search, event investigation, evidence extraction, and automated alerts. ### Core Features 1. Multi-stream video ingestion and preprocessing. 2. AI-powered object detection and segmentation. 3. Object tracking and movement analysis. 4. Zone-based monitoring and event detection. 5. Visual and text-based semantic search. 6. Natural-language video querying. 7. Multimodal intelligence and contextual reasoning. 8. Timestamped evidence generation. 9. Anomaly detection and automated alerts. 10. Centralized multi-camera monitoring. ### Technology Stack - **AI Models:** SAM 3.1, SigLIP2, BGE-M3, Muse Spark 1.3. - **AI Inference:** Baseten (three independent model deployments). - **Backend:** Python, FastAPI, Uvicorn, Pydantic. - **Video Processing:** OpenCV, FFmpeg, FFprobe. - **Vector Database:** Qdrant. - **Database:** PostgreSQL, SQLite. - **Object Storage:** MinIO / S3. - **Background Processing:** Redis, Celery. - **Frontend:** HTML, CSS, JavaScript. - **Deployment:** Docker, Docker Compose. ### Project Objective To develop an intelligent video intelligence platform that transforms raw surveillance footage into structured, searchable, and actionable information, enabling users to investigate video events through AI-powered visual understanding, semantic retrieval, and natural-language interaction.