sunilmehta.si · Projects · Experience · Resume · LinkedIn · X · Email
I'm Sunil Mehta, Founding Lead DevOps & MLOps Engineer at ORGN, based in Gurugram and open to remote roles. For 9+ years I have built and run the platforms AI products depend on, from multi-cloud Kubernetes and confidential GPU compute to secure delivery and cost governance.
In the open I build benchmarks and tools that answer questions a platform team has to answer, and I send the fixes I find upstream.
| Project | What it asks | What held |
|---|---|---|
k8s-gpu-scheduler-benchmarkv0.5.0 |
Which Kubernetes GPU scheduler should a platform run? | Out of the box only KAI bin-packs GPU pods by their GPUs. One setting made Volcano 26% faster and 26% cheaper per useful GPU-hour on the same trace. |
k8s-inference-routing-benchmarkv0.6.0 |
How much does the router in front of LLM servers matter? | Prefix-aware routing lifts cache hits from 64% to 80% and cuts median time to first token by 27%. Under load the router matters more and the routers differ less. |
k8s-inference-autoscaling-benchmarkv0.8.0 |
What should an autoscaler follow when it scales LLM inference? | On a real cluster no ramp from 30 seconds to 10 minutes was slow enough for the default HPA to keep requests on time. Pod startup is the constraint, a small warm pool is the cheapest fix, and a forecast helps on slow ramps only. |
k8s-gpu-cost-scannerv0.1.0 |
Which GPUs is the cluster paying for and not using? | A read-only Go tool on the metrics a cluster already has. On a synthetic fleet backfilled into a real Prometheus in CI it finds every planted problem: 221.4 wasted GPU-hours in a day, each counted once, and the cost per million tokens of each model. |
kube-diagnosev0.5.0 |
Why is this Kubernetes workload broken? | A read-only kubectl plugin that names the root cause with the evidence and the kubectl command behind each fact. On a real cluster in CI it named the cause of 87 of 87 broken namespaces in its first finding; k8sgpt named 33. record --redact makes a snapshot safe to share. |
kubernetes-secure-mlops-pipelinev0.8.0 |
Can every model in production be reproducible, signed and verified? | A real kind cluster runs 24 checks on every push, and the gate blocked a 0.909 challenger against the 0.938 champion. |
agentic-code-review-toolv0.11.1 |
Scanner-grounded AI review for the whole pull request. | 128 rules, each swept over real repositories before release. One rule went from 1,556 findings to 101. |
Also: ai-devsecops-cicd, a secure delivery path for LLM and agent services · llm-inference-platform, LLM serving on Apple Silicon you can measure · jev-inference-router, an observable routing layer for local inference.
- One workload, every contender. The same trace is replayed through each scheduler, router or policy, so a difference belongs to the component and not to the traffic.
- Real components, in CI. Kueue, Volcano, KAI, the llm-d router, the HPA and KEDA run as released, on kind clusters in GitHub Actions. Anyone can run them again.
- Every number has a test. The datasets are committed, and tests hold each figure a README quotes to the data behind it.
- Only what repeats is claimed. A difference on a cluster counts when it holds over more than one run. One early reading reversed on the second run and was withdrawn in the open.
| When | What | |
|---|---|---|
| 17 Nov 2026 | kube-diagnose v0.5.0 |
13 new rules, from volumes a container may not use to pods on the node's DNS and CronJobs that skip runs: 87 of 87 faults named first on a real cluster; k8sgpt named 33. |
| 12 Nov 2026 | k8s-gpu-cost-scanner v0.1.0 |
A new tool: the GPUs an inference cluster pays for and does not use, and what each model costs per million tokens. |
| 11 Nov 2026 | k8s-inference-autoscaling-benchmark v0.8.0 |
A predictive policy: scaling one pod startup ahead helps on slow ramps, and warm capacity still wins the fast ones. |
| 09 Nov 2026 | kubernetes-secure-mlops-pipeline v0.8.0 |
Every pin is held back until its release is 14 days old, and reported when it drifts. |
| 04 Nov 2026 | agentic-code-review-tool v0.11.1 |
A false positive found by sweeping Inspect AI: 1,124 findings from one rule, down to 8. |
| Ship | Guard | Run | Watch & own |
|---|---|---|---|
| GitHub Actions, keyless through Workload Identity Federation | Secret, code, dependency, image and infrastructure-code scanning | 10+ GKE clusters for an engineering org of 50+ | 99.95% uptime across 200+ services |
| Artifact Registry, images signed with cosign | Kyverno and OPA at admission | 300+ Intel TDX confidential GPU nodes | Istio, Prometheus, Loki, OpenTelemetry, Grafana HA |
| Argo CD to every cluster | Falco at runtime | Rust LLM gateways, 100M+ requests a month | $70K–$90K of monthly cloud spend owned |
| Protected Terraform workflows | OAuth 2.1 / OIDC sign-on, migrated with zero downtime | Infrastructure cost cut by 55% |
This describes my professional work. Most of it lives in private organization repositories. The full experience, five roles since 2017, is on the site.
| Project | Contribution | Status |
|---|---|---|
| Argo CD | Documented OCI repositories over plain HTTP (insecureOCIForceHttp, type: oci secrets) |
Merged |
| MLflow | Model downloads no longer modify signed model directories, which broke signature verification. Reported, then fixed. | Merged |
| NVIDIA KAI Scheduler | Corrected the GPU fractioning docs to the names the subchart and operator use | Merged |
| llm-d inference simulator | Render requests no longer fail with EOF when vLLM closes an idle connection. Reported, then fixed. |
Merged |
| KServe | Security: every Python lock moved off the certifi that still trusted the compromised emSign roots | In review |
| Tesla vehicle-command | Fixed key storage rejecting valid P-256 keys whose scalar starts with a zero byte, about one key in 256 | In review |
| xAI Python SDK | Shipped the missing py.typed marker, so type checkers use the SDK's annotations |
In review |
| X API CLI (xurl) | The webhook server fails fast without an ngrok token instead of hanging | In review |
| etcd-operator | The operator no longer carries on without its TLS certificates when provisioning fails | In review |
| KServe website | Install docs: create the kserve namespace before applying the manifest |
In review |
| Kubeflow Pipelines | A compiler bug: a multi-line default value makes the compiled pipeline YAML invalid. Fix ready. | Reported |
Infrastructure Kubernetes · Terraform · Argo CD · Helm · GCP · AWS · DigitalOcean · Cloudflare
MLOps & serving Kubeflow Pipelines · MLflow · KServe · vLLM
GPU scheduling Kueue · Volcano · KAI Scheduler · kwok
Security Intel TDX · OAuth 2.1 / OIDC · Kyverno · OPA · cosign · SLSA · OpenSSF model signing
Observability Prometheus · Grafana · Loki · Tempo · OpenTelemetry · ClickHouse
Languages & data Go · Rust · TypeScript · Python · PostgreSQL · Redis
Building AI platforms, developer tools or secure infrastructure?
sunilmehta.si · LinkedIn · X · Email

