Deep dives into cloud architecture, DevOps practices, AI engineering, and the journey to senior technical leadership.
How I built an AI governance platform on AWS EKS — FastAPI inference endpoint with per-request audit logging, model card endpoint, fairness metadata, and Helm-packaged deployment with HPA.
How I built an end-to-end MLOps pipeline with SageMaker Pipelines, automated retraining via EventBridge, and drift monitoring using KS tests and CloudWatch — for a telecom churn use case.
Learn how to build a telecom customer churn predictor using Random Forest, Keras neural networks, and deploy it to a real-time SageMaker endpoint. Full code included.
How I built a real-time IoT data pipeline on AWS — device simulator → Kinesis stream → Lambda consumer → DynamoDB — with anomaly detection that fires SNS alerts and CloudWatch metrics.
The industry ships agents fast and debugs them in production. Here's the opposite approach — local-first agentic development, liftability by design, and selective promotion to AWS Bedrock AgentCore.
Most AI projects never reach production. The missing piece is prompt testing — with the same rigour as TDD. Here's the strategy for shipping AI systems that actually work.
Our first LLM judge gave a 9/10 to a page with invisible text. Here's how we fixed it with structural guardrails, multimodal inputs, and a fixed-weight violation catalogue.
Most LLM benchmarks evaluate text. We needed to evaluate entire websites. Here's the 4-layer evaluation framework we built to score AI-generated multi-file artifacts using a violation-deduction model.
We tested Claude Sonnet, Kimi K2.5, Claude Haiku, DeepSeek V3.2, and DeepSeek R1 on the same 16-action website generation pipeline. The results weren't what we expected.
A practical guide to deploying a local Python MCP server to Amazon Bedrock AgentCore Runtime — from localhost prototype to production-grade cloud service with session isolation, authentication, and observability.
A six-layer memory architecture for persistent AI agent knowledge — from CLAUDE.md foundations to auto memory, plans, and permissions — managing 14 specialized agents across multiple AWS projects.