My Projects

Work from the job — Funavry client platforms. Start with the EKS migration if you only read one.

Start hereClient · ChainGPT

Legacy Platform → Consolidated EKS

Scattered ECS, EC2, Lambda, and per-account stacks onto two GitOps-managed clusters.

Problem

Workloads lived across AWS accounts — ECS, EC2, Lambda, extra ALBs, no shared GitOps.

Owned

Target platform, Terraform EKS, Istio, Karpenter, and the production cut-over.

Result

Two GitOps clusters; private-by-default APIs; fewer load balancers and log forwarders.

AWS EKSTerraformArgo CDIstioKarpenterExternal Secrets

Read write-up

Client · Normies & ChainGPT

Reusable GitHub Actions Deploy Platform

One pipeline for every service — security gated, health-checked, auto-rollback, and notified.

Problem

Every repo invented its own deploy YAML. Green CI meant the image pushed, not that the cluster was healthy.

Owned

Reusable workflow_call suite: OIDC, scan, GitOps bump, health gate, auto-rollback, notify.

Result

New service is config. CI green means the cluster is healthy — and a failed gate rolls itself back.

GitHub ActionsOIDCECRTrivySemgrepArgo CD

Read write-up

Client · ChainGPT

Dotfun Agents Platforms

Multi-tenant agent factory on EKS — GitOps provision, in-cluster mesh, DNS at the edge.

Problem

Thousands of tenants only works if create and destroy are APIs, not git commits and DNS tickets.

Owned

Path from activate to a healthy public URL — Argo CD as an API, Helm, ExternalDNS, Cloudflare.

Result

Isolated agents with their own hostname; teardown prunes compute and DNS.

AWS EKSArgo CDHelmKustomizeExternalDNSCloudflare

Read write-up

Client · ChainGPT

Private IPFS Cluster on Wasabi S3

Terabytes of blocks on object storage, not SSDs.

Problem

A disk-only replica holding millions of pins used about 6–7TB locally — expensive to scale and recover.

Owned

Private IPFS Cluster with hybrid disk + Wasabi S3 block store, GC, and pin-set ops.

Result

S3 replicas stay ~100GB local; pins stay in sync across regions without another 7TB volume.

IPFS KuboIPFS Clustergo-ds-s3DockerWasabi S3LevelDB

Read write-up

Client · ChainGPT

Self-Hosted BSC Full Node (Internal RPC)

In-house BNB Smart Chain RPC — cheaper than paid providers, keyed, rate-limited, and actually monitored.

Problem

Backends sat on paid public RPC. The in-house node was stale, paging falsely, and had no API keys.

Owned

Snapshot restore, pruneancient, keyed gateway, and alerts that check eth_syncing first.

Result

Internal RPC on the hot path; noisy clients get 429; pages match real lag vs chain tip.

UbuntuDocker ComposeBSC GethPBSSJSON-RPCHAProxy

Read write-up