Challenges Solved
- Model Consolidation: Developed a unified "hydra-net" model consolidating 3 separate models into 1, reducing memory usage by 50%.
- Inference Optimization: Deployed PyTorch inference on Azure GPU with autoscaling, cutting vendor API costs by ~60%.
Signal
Computer Vision / Cloud Cost Optimization
Project Details
Developed as a mission-critical component for secure document verification at ARGO DATA. The system performs classification, data extraction (OCR), and forgery detection on various ID documents.
Key Innovations
- Hydra-Net: A custom-designed neural network with three task-specific heads (Classification, Alignment, Quality Assessment) sharing a single feature extraction backbone.
- In-House OCR: Optimized PyTorch-based OCR pipeline tailored for the specific fonts and layouts found in government IDs.
Technical Depth
- Optimized Backbones: Used lightweight but powerful backbones for the Hydra-Net architecture to enable fast inference on CPUs and small GPUs.
- Autoscaling Deployment: Integrated the system with Azure Kubernetes Service (AKS) to automatically scale workers based on document upload volume.
- Data Privacy: Built the system to operate entirely within the secure ARGO DATA cloud perimeter, ensuring PII is never exposed to external vendors.
Links
- Internal Project (ARGO DATA)