Skip to content

// what I work on

What I do

I build the platforms other engineers ship on, and the infrastructure underneath them. These are the areas I go deep in: ML platforms, GPU, Kubernetes, platform security and observability — the systems that keep production honest.

01

ML Platforms & MLOps — Zero to Production

ML platforms built from scratch for several teams at once: notebooks and pipelines (Kubeflow), experiment tracking and model registry (MLflow), model serving (KServe), LLM observability (Langfuse) and data labeling (Label Studio), fed from Snowflake and S3. Deployments that accept only registered, versioned models, and per-team workspaces with SSO and Vault-issued credentials.

02

GPU Infrastructure

Shared GPU capacity instead of one card per person: fractional H100 sharing with selectable VRAM slices, several GPUs per workload when needed, quotas checked at admission that explain a refusal before launch, and GPU nodes provisioned on demand.

03

Developer & AI Tooling

SDKs, command-line tools and MCP servers that give people and AI assistants one authenticated way into a platform from a laptop — no kubeconfig, no cloud keys, and state-changing operations behind an explicit opt-in. Plus LLM-backed applications with retrieval and memory behind them.

04

Kubernetes & Cloud Infrastructure

Production Kubernetes on AWS (EKS) and bare-metal, built with OpenTofu/Terraform and delivered with Argo CD — multi-region clusters, Istio service mesh with mTLS, Kyverno policy enforcement, autoscaling, and requests sized to real usage. Reproducible, reviewable, GitOps-driven.

05

Platform Security & Secrets

Defense-in-depth: HashiCorp Vault with internal PKI and short-lived credentials, SSO/OIDC (Keycloak / Authentik / Pocket ID), Kyverno and Pod Security Standards, image scanning (Trivy) in CI, intrusion prevention, network policy, mTLS, and tenant isolation done properly.

06

Observability & Reliability Engineering

Prometheus / VictoriaMetrics and Grafana stacks that scale to high-cardinality workloads, OpenTelemetry distributed tracing, meaningful SLOs, alerting that doesn't cry wolf, and dashboards that shorten incidents.

07

Self-Hosted & On-Prem Infrastructure

Proxmox and Kubernetes clusters defined entirely as code, private registries (Harbor/Nexus), self-hosted GitLab with CI runners on Linux, Windows and macOS, and GPU/NPU passthrough for local AI workloads. Sovereignty without the chaos.

// how I work

How I work

I build systems meant to be read: documented, reproducible, and reviewable — not black boxes. Decisions get written down as ADRs, infrastructure lives in code, and when an upstream project gets something wrong, the fix goes back upstream.