Milind Dethe

Building boring, reliable infrastructure
for the cloud-native world.

edu
IIT KGP '26
prev
Atlan · platform
oss
LFX '24 Thanos · Kubeflow
cert
CKA · Aug 2026

platform · devops · kubernetes · ai infra

resume · email · github · linkedin · x/twitter

experience

Jun — Nov 2025

Atlan — Software Engineer Intern, Platform

Automated EKS control-plane and node upgrades with Argo Workflows (80% faster), synced secrets with External Secrets Operator, built Grafana dashboards and alerts that cut MTTD by 40%, and served on the platform on-call rotation.

Jan 2025 — Now

Kubeflow — Contributor · CNCF

Cascade deletion in the Pipelines backend, release automation for Kubeflow Trainer, and Release Shadow for Kubeflow Community Distribution 26.03.

Sep — Nov 2024

CNCF Thanos — LFX Mentee

Optimized request hedging in Thanos Store Gateway with adaptive T-Digest thresholds, cutting P99 tail latency by 30% — merged upstream as an opt-in production option.

projects

see all
LLM Inference Platform kubernetes · kserve · vllm · argo cd · helm A production-style LLM inference platform built with KServe, vLLM, llm-d, Envoy Gateway, Argo CD, and Helm — exploring scalable model serving, request routing, and GitOps workflows. DeployGuard go · kubernetes · prometheus · operator sdk A Kubernetes operator that analyzes Prometheus metrics during rollouts to automatically detect regressions and enable safer, metrics-driven deployments. Predictive Kubernetes Autoscaler keda · python · fastapi · kubernetes · prophet An intelligent autoscaling system that forecasts workload demand and proactively scales Kubernetes applications using KEDA and time-series predictions. Store Provisioning Platform argo workflows · kubernetes · aws · go · gitops Automated provisioning of isolated MedusaJS stores using Argo Workflows, enabling repeatable deployments with GitOps and cloud-native infrastructure.

open source

61 merged upstream
store: support hedged requests thanos-io/thanos fix: delete a pipeline along with all its versions kubeflow/pipelines feat: run CI workflows on /ok-to-test label kubeflow/trainer receive: unhide tsdb out-of-order flags thanos-io/thanos fix: respect browser theme preference on initial load prometheus/docs

writing

see all
Day 7: Deployment Is Not The End Day 6: Why Do Models Fail Even When Training Accuracy Looks Great? Day 5: When Does a Training Script Become a Pipeline? Day 4: Training One Model Is Easy. Finding The Best One Is Hard. Day 3: If Git Versions Code, What Versions Data? Day 2: From Notebook to Reproducible Training