Building secure cloud systems today and intelligent AI systems for tomorrow
I'm Praneeth Kodumagulla, a full-stack platform engineer at HSBC and an independent researcher exploring AI agents and automation. My engineering work spans cloud-native applications, Kubernetes, CI/CD, infrastructure automation, and observability, with a focus on reliable systems in financial services. Alongside that work, I build agent prototypes and study how instructions shape autonomous systems’ behavior. I bring a platform engineer’s perspective to AI: how do we test these systems, understand their failures, and make their behavior more dependable? I share what I learn through hands-on tutorials, reproducible experiments, and practical code.
Markdown supported · ⌘↵ to post
Really thorough methodology, especially separating the client-side watch timeout from the controller's own ProgressDeadlineExceeded, that distinction trips people up constantly. One thing worth asking about the recovery step: reapplying the same healthy spec created a brand new revision each time (2 to 4, 4 to 6, 6 to 8) instead of landing back on an existing one. In a setup where you're triggering an automated rollback off ProgressDeadlineExceeded, does kubectl rollout undo without an explicit --to-revision reliably land on the last genuinely healthy spec, or does it just walk back to the most recent distinct ReplicaSet regardless of whether that one was itself a different failure? Two failures in a row would make that ambiguity matter a lot more than the single-fault case this lab walks through.