What we’ve learned shipping production AI, web, cloud and software — written by the engineers doing the work, not a marketing team.
Six months in, here’s what worked: small tools, narrow scopes, real evals, humans in the loop where the cost of being wrong is high — and the patterns we threw away.
Read storyHybrid retrieval, citation-required prompts, refusal-when-uncertain, and a real eval harness — the four levers that actually moved the metric.
Read moreAn offline-first mobile app, a clean cloud sync layer and a small AI module that kills duplicate work. 72% reduction in admin overhead in three months.
Read moreA 90-day cost review walkthrough: what we measured, what we right-sized, the surprises (savings plans aren’t always the answer) and the playbook we now reuse.
Read moreGitHub Actions + Terraform + an opinionated test layer. Boring? Yes. But it ships, it’s observable, and onboarding a new repo takes an afternoon, not a week.
Read moreMost “backup” programmes never test the restore. We walk through our monthly DR drill, the runbook we use, and the surprises we found in our own production.
Read moreA practical decision matrix from our automation work: cost of being wrong, reversibility, audit needs, and the sneaky cases where humans are the bottleneck.
Read moreFederate identity through Azure AD, broker access through AWS Identity Center, mint JIT database credentials — and retire long-lived access keys for good.
Read moreIngest GitLab + Kubernetes + Splunk through typed MCP tools. Classify the failure. Re-run, draft an MR, or escalate — with a hard wall between the model and your main branch.
Read moreStitching Prometheus, CloudWatch, GuardDuty and CloudTrail into a single Grafana view that engineers, SRE and the business all read from — and what we deliberately left out.
Read moreNo fluff, no listicles, no “10 ways AI will change everything.” Just what we shipped, what broke, and what we learned. Unsubscribe any time.