John LeeJuan Lee

Founded Sigmoid to make cloud infrastructure easier to operate. I carry the product end to end, from planning and design through the front end, the back end, and the execution infrastructure that runs customer deployments. I built and piloted a tool that compiles an architecture drawn on a canvas into Terraform and deploys it to AWS, and I'm redefining the product direction based on how the pilot teams actually operate their infrastructure.

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 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 setup hit the limits of VPC Peering and NLB, and adopting Istio then added Sidecar resource usage and operational load on top.
• Moved to a Service Networking architecture where VPC Lattice handles cross-cluster connectivity and Cilium handles network policy and traffic observability, and resolved the connectivity and bootstrap issues that came up during the migration.
• Saved 76.8 vCPU and 150 GiB of memory per 300 Pods.

• Identified a shared pain point across many startup meetings, that AI had made writing code easy but deploying and operating it was still hard, and designed a Visual Infrastructure Platform that compiles infrastructure drawn on a canvas into Terraform with stable IDs and a reference structure, then deploys it.
• Received pre-registrations from companies including Coupang and Deloitte.
• Confirmed through the pilot and user interviews that hiding infrastructure isn't enough and teams need to understand what they run, and am redesigning the product to help them do that.

• Developed an open-source chaos engineering platform for Kubernetes, structured around an Operator, Executor, and Daemon, to verify cluster behavior under failure repeatedly. It ships with a CLI and Helm charts and implements 20 fault experiments across Pod, Network, Node, and Stress.
• Implemented Pod-level faults that trace down to the host-side veth and cgroup v2 path, so network and resource faults apply only to the targeted Pod instead of the whole node.
• Limited each run's blast radius with BlastRadiusPolicy and implemented an Abort Condition that halts the run as soon as a steady-state probe drifts out of bounds.
• Designed recovery to restore a snapshot of the pre-fault state instead of guessing at the reverse command, so the outcome of a rollback is predictable.

• Normalized heterogeneous detection events from GuardDuty, Falco, and Tetragon into a common schema, deduplicated them, ranked them P1-P4, and designed a four-stage agent chain (Summary → Triage → Solution → Remediation) to analyze them.
• Applied Claude Haiku to Summary and Triage and Claude Sonnet to Solution and Remediation, where the response is generated and the execution decision is made, tuning cost and response speed by stage.
• Implemented automatic remediation for P3-P4 events and Slack approval before P1-P2 execution, and logged every analysis result, approval, and execution to DynamoDB as an audit trail.
• Routed Kubernetes changes through a permission-limited EKS MCP server instead of letting the agent run them directly, and implemented a Destructive Action Guard in the execution layer that prevents the agent from executing destructive actions until Pod snapshotting and Tetragon SIGKILL labeling are complete.

AI Research and Development Club; PARA