Co-owner of the production RPC platform: node fleets, routing, observability, GitOps, on-call.

• scaleOperate 50+ blockchain node fleets (Polygon, Ethereum, Bitcoin, Solana, BSC, Cosmos) on Kubernetes and bare metal, full lifecycle from rollout to incident response.
• routingContributed to moving platform-API and chain-RPC admission from nginx-Lua to Envoy Gateway (Gateway API, extAuth) with the SLO budget intact through cutover.
• gitopsAuthored the progressive multi-region rollout patterns on ArgoCD ApplicationSets, and the runbooks for their failure modes.
• discoveryDesigned unified service discovery (Consul) across Kubernetes, Nomad, and bare metal: 60 members, 200+ RPC/WS endpoints.
• upstreamChase production bugs down into the chain clients themselves and fix them upstream: Erigon PR #22877 (non-deterministic eth_estimateGas, a cancellation race on a shared EVM).
• toolsBuilt the internal tooling layer: vitals snapshot CLI, dashboards-as-code, ZFS-snapshot rollback for chain upgrades, MR-review knowledge base.

Sole DevOps engineer on an AWS-native stack. Took the platform from no CI and no monitoring to a working IaC and observability foundation.

• ci/cdHA Jenkins on Kubernetes via JCasC; Gitflow, peer review, and IaC introduced as engineering practice.
• o11yPrometheus, Grafana, and Loki: the first real observability the team had.

Short-term cloud and DevOps contract engagements.

• chaosResiliency testing with chaos-toolkit; Consul KV sidecar POC for hot-reload configuration.
• k8sPython tooling against the Kubernetes API for cluster-component health verification on deploy.

Part of the team that introduced DevOps practices org-wide: containerized legacy Java applications onto Kubernetes, built the shared Jenkins pipeline platform, and owned the CI/CD migrations end to end.

Built ezmarc.com with a teammate: a Python/Flask API converting book metadata to MARC records.

A watcher goroutine could cancel the shared EVM after its own probe finished, aborting the next probe with err == nil, so the gas binary search converged below the true minimum. Replaced it with context.AfterFunc plus a joined stop(); added a 200-call determinism test. Writeup →

The sparse-attention decode gate excluded sm_121, so GB10 fell to a fallback kernel that aborted at CUDA graph capture and the model could not serve. Widened the gate; verified on 2x DGX Spark serving Qwen3.8-Flash-Next. Writeup →

View certificate →