A walkthrough of FinOps for agentic platforms — where the money actually leaks, why provisioned concurrency isn't always right, and a battle-tested pattern for keeping AI compute under 30% of revenue.
Read the essay →Side-by-side benchmark and operational notes from running both in production. Recall, ingest throughput, hybrid search, and the operational tax nobody talks about…
What I'd do differently if I were starting Ripio's 6-country topology again. DynamoDB Global Tables, Aurora Global, Route 53 health checks, and where the latency math actually bites…
API Gateway → Lambda → DynamoDB single-table + Step Functions + EventBridge. Why provisioned concurrency mattered, where Step Functions hurt, and the single-table schema that…
What I look for after 30+ years and three founding teams. Hint: it's not the cert list.
Today's reminder: spot price is not a strategy. Spot diversification is.
If your post-mortem template doesn't have a 'systemic' section, it's a blame template.
Bedrock + Claude for tool-use, OpenAI for streaming-heavy chat, Anthropic API direct for evals. Right tool for the job.
Karpenter + spot is the single biggest lever for EKS FinOps. Nothing else is close.
32 years in IT and I'm still rebuilding my dotfiles every Sunday morning. Some things never change.
Opinionated EKS + Karpenter module battle-tested at 6-country scale.
CDK construct library for production RAG on AWS (OpenSearch + Bedrock + Lambda).
Blameless post-mortem template I've used across 4 companies.