Skip to main content

🦅 Alert.CA.CONDOR

Overview​

CONDOR (Computational Observation for Natural hazard Detection and Optimized Response) is a research-only computer-vision model developed in collaboration with UC San Diego (UCSD) to explore whether a machine learning model could identify early wildfire indicators in fixed-camera imagery collected by the UCSD-operated Alert California program.

Key Facts
AspectDetail
PurposeResearch & experimentation only — not production or safety-critical
InputStill images from UCSD-operated fixed cameras
OutputProbabilistic indicator of wildfire-like visual patterns
CollaborationMicrosoft AI for Good Lab × UCSD Alert California

System Components​

ComponentDescription
ClientPolls camera APIs, runs inference via Triton, and pushes events to the server
ServerFastAPI backend with PostgreSQL, Redis, analytics, export, and a review dashboard
Model RepositoryTensorRT-optimized smoke/fire detection model served by Triton Inference Server
DatabasePostgreSQL for events/annotations, Redis for sessions/analytics/caching
Data ExportStreaming COCO dataset generation for ML training pipelines
InfrastructureAzure GPU VM + Container App, deployed via Bicep orchestration

Architecture​

Documentation Structure​

  • Analytics — What we track, API endpoints, and how to read reports
  • Authentication — Entra ID SSO, session management, CSRF, and API keys
  • Deployment — End-to-end Azure deployment guide
  • Export — Streaming COCO dataset generation for ML training
  • Model Deployment — Deploying a new model version (ONNX → TensorRT)
  • Components — Per-component docs (Client, Server, Dashboard UI, Redis, Scripts)
  • Runbook — Operational procedures for infrastructure and troubleshooting