Ideas2IT engineered an AI platform that detects risks in residential utility footage using an ensemble of object detection models. We trained, optimized, and deployed the solution with CI/CD pipelines built for high performance and zero disruptions.

Core Implementation Highlights:

• Improved Model Precision and Recall
  • Class imbalance was addressed via augmentation of rare utility types.
  • Background image noise was added to reduce false positives in inference.
  • A curated dataset was created from real annotated images, filtered by bounding box size and quality.
• Ensembled YOLOv5 Architecture
  • Multiple YOLOv5 models were trained and ensembled to maximize detection accuracy without sacrificing inference speed.
  • Test-Time Augmentation (TTA) was applied for high-confidence detection in edge cases.
• Zero-Downtime Deployment Framework
  • Containerized models with Docker and deployed using Azure Kubernetes Service (AKS).
  • CI/CD pipelines supported horizontal and vertical scaling, with zero lag during model upgrades.
• Custom Preprocessing Layer
  • Built on Azure Functions to handle:
    • Deduplication
    • DOB normalization
    • Standardization of names and phone formats
  • Ensured clean, consistent metadata for every inspection submission.