Tractable

Key Contributions

• I anonymised 25 million image assets with a pipeline I implemented as the groundwork for our computer vision model ingestion pipeline. Our design estimated a timeline of approximately one week to complete but the final implementation was complete within 1.5 days, a 22% performance improvement (and cost saving for the customer). The anonymisation models were loaded into an AWS Lambdas, with our custom code deployed onto the customer's AWS ('on-prem'). I implemented the infrastructure as code, which was audited by external security experts and included private VPCs, KMS secured asset transfer. The core logic handled JPEG and PNG image ingestion, PII detection (panel, text, faces) via PyTorch models with custom segmentation, and blurring validated ROIs using irreversible kernels.
• Built secure AWS-native deployments using IaC (VPC isolation, KMS, Terraform/CDK) audited by external security consultants.
• Built backend proxy and enrichment layers for event routing, fraud-detection workflows and secure callback processing using API Gateway and Lambda.
• Implemented stateless architectures and secure secrets handling using SSM and parameterised infrastructure.
• Deployed observability & reliability patterns including Grafana dashboards, incident response and reprocessing flows.