Sai Ram

AI Infrastructure and Platform Architect · Chennai, India

I help engineering teams measure what each AI workload costs per successful task and compare the options against their quality and latency targets: a managed model API, their own GPUs, or a mix. I have 15+ years of experience in infrastructure and platform engineering. I hold CKA and CKS and contribute to KEDA, the Kubernetes autoscaler.

Try it

AI inference cost explorer: hosted model API, your own GPU, or a mix? Set your volume and constraints and see what each option costs per correct answer, from measured runs.

Sample results dashboard: evidence from my own benchmark runs, covering cost, routing, service targets, failure drills and autoscaling.

Recent work

Inference Cost Optimization is a reproducible, open case study for deciding when a managed model API, a private vLLM service, or a policy-routed hybrid is the lower-cost choice after quality, latency, security, and reliability constraints are applied. Eight published runs so far, measured with Prometheus, Grafana, and DCGM, including autoscaling under load with KEDA. Every number in it carries a date, a source, and a comparison boundary.

Notes

Where 7m 40s of vLLM startup went: one KEDA scale-up, second by second, from the run's events and pod status, and how to check your own cluster.

Tools

inferaudit is a read-only review of how AI models are served on a Kubernetes cluster: startup timelines from Kubernetes events and pod status, plus reliability and access checks on the configuration, in one HTML report. Here's the sample report, built from a reconstructed snapshot of my lab.

Work with me

The AI workload cost assessment takes five working days on one workload and ends with a written recommendation, which may be to keep your current setup. I also take on serving stack reviews and platform builds. Write to me with a line about the workload and what it runs on today.