John Lee

Founded Sigmoid to make cloud infrastructure easier to operate. First built an AI-based IaC generation and infrastructure management platform (2025.10–2026.02), then pivoted (2026.03). Now building a tool where you draw cloud infrastructure on a canvas and it deploys straight to AWS, leading design and backend so teams can run their own cloud without dedicated DevOps/Infra engineers (2026.04–present).

Designed a multi-cluster EKS networking architecture with Amazon VPC Lattice and Cilium ENI mode, replacing a sidecar-based mesh's operational overhead and resource cost. VPC Lattice handles cross-cluster connectivity, while Cilium's eBPF data plane provides network policy and traffic observability. Saved 76.8 vCPU and 150 GiB per 300 Pods; presented the work at AWS Summit Seoul 2026 (DEV305).

AWS Summit Seoul 2026 - DEV305

As per-customer infrastructure and data isolation requirements grew, the EKS cluster setup hit the limits of VPC Peering and NLB. After adopting Istio, Sidecar resource usage and operational load increased, so I moved to a VPC Lattice and Cilium based Service Networking architecture, resolving connectivity and bootstrap issues. The result saved 76.8 vCPU and 150 GiB of memory per 300 Pods, cutting roughly $42,000 per year.

Across many startup meetings I found a shared pain point: AI made writing code easy, but deploying and operating it was still hard. I designed and founded a Visual Infrastructure Platform where you draw infrastructure on a canvas and it auto-compiles and deploys to Terraform with stable IDs and a reference structure. It secured pre-registrations from companies including Coupang and Deloitte.

Built ChaosPlane, an open-source chaos engineering platform for Kubernetes, structured around an Operator, Executor, and Daemon so I could repeatedly verify how the cluster behaves under failure. It's organized into 6 repositories, including a CLI and Helm charts, and supports 20 fault experiments across Pod, Network, Node, and Stress categories. Pod-level faults trace down to their host-side veth and cgroup v2 path, so network and resource faults land on the Pod I target, not the whole node. Blast radius is capped with BlastRadiusPolicy, and an Abort Condition immediately halts a run the moment a steady-state probe drifts out of bounds. Recovery doesn't guess at the reverse command. Instead, it snapshots the pre-fault state and restores it as-is, so the outcome of a rollback is predictable.

In EKS, I normalized heterogeneous detection events from GuardDuty, Falco, and Tetragon into a common schema, deduplicated them, and ran them through a four-stage Agent Chain: Summary → Triage → Solution → Remediation. Events are classified into a P1–P4 priority; Claude Haiku handles Summary and Triage, while Claude Sonnet — which generates the response and decides whether to execute it — handles Solution and Remediation, tuning cost and response speed by stage. P3/P4 events get remediated automatically, while P1/P2 only execute after Slack approval, with every analysis result, approval, and execution logged to DynamoDB as an audit trail. Kubernetes changes are never run by the agent directly — they go through an EKS MCP server with limited permissions, and a Destructive Action Guard in the execution layer blocks destructive actions until Pod snapshotting and Tetragon SIGKILL-labeling are complete.

AI Research and Development Club; PARA