Engineering Blog
Technical deep-dives and architectural decisions by the Nom AI team.
Scaling Arabic OCR to 10M Pages/Day using Amazon EKS
Cloud Architecture • Sept 2026
When our largest banking client required us to process 10 million historical documents over a weekend, our standard EC2 auto-scaling groups weren't responsive enough. In this post, we detail our migration to Amazon EKS using Karpenter for aggressive node provisioning, combined with AWS Inferentia (Inf2) instances to drastically reduce our cost-per-inference.
Implementing Zero-Trust VPC Peering for Financial Institutions
Security & Compliance • Aug 2026
Banks cannot send sensitive KYC documents over the public internet. Discover how we architected a Zero-Trust data plane using AWS PrivateLink and cross-account IAM roles, allowing financial institutions to consume our AI models securely from within their own isolated AWS environments.
Optimizing Transformer Models for Right-to-Left (RTL) Languages
Machine Learning • Jul 2026
Standard OCR engines fail miserably at cursive Arabic due to bounding box overlaps. We explain our custom attention mechanism modifications in PyTorch and how we achieved a 99.4% accuracy rate on blurred, handwritten Saudi national IDs.