Hyunsuk Bang

• Automated lifecycle management of Kubernetes clusters running NVIDIA and AMD GPUs, including health monitoring and auto-remediation of unhealthy nodes.
• Contributed to Cilium (CNCF) — fixed a race condition in the IPSec key rotation path that caused packet drops during agent rolling restarts. Solution used pinned BPF maps to defer peer advertisement until kernel XFRM states were ready (PR #44701).
• Drove the migration from kube-proxy to Cilium's kube-proxy-replacement across all clusters, auditing the Cilium codebase to validate impact and correctness of service routing.
• Led the migration of Kubernetes clusters from Ingress NGINX to Gateway API + Envoy.
• Enabled secure arbitrary code execution for AI workloads by architecting a sandboxed runtime on Kata Containers backed by Firecracker microVMs — VM-level isolation inside Kubernetes for running untrusted code.

• Migrated computationally intensive Kubernetes clusters from AWS to an on-prem data center, achieving up to 90% cost savings.
• Built a tool to query Mimir and Prometheus across 300+ Kubernetes clusters, identifying over- and under-utilized workloads with automated email notifications for owners.
• Analyzed metric cardinality in Mimir to achieve a 10% reduction in cluster workload, and built a Grafana dashboard for ongoing high-cardinality monitoring.
• Streamlined Cilium maintenance on ArgoCD and authored technical documentation and best practices.
• Used Prometheus, Grafana, OpenTelemetry, and Tempo to collect, visualize, and analyze metrics and traces for faster incident resolution.

• Conducted in-depth research and built comprehensive test plans for evaluating the performance and accuracy of a BPF compiler.
• Designed packet-processing applications hitting 10 Gbps in-kernel and 100 Gbps on DPDK using Receive Side Scaling and parallel processing.
• Built automation scripts to ensure replicability of test results across runs.
• Collaborated with Fermi Lab on a joint project that leveraged DPDK to generate measurement output.
• Documented findings, methodologies, and procedures in technical reports and presentations.

• Built a compiler that translates pcap expressions into Berkeley Packet Filters for more efficient and accurate packet filtering.
• Designed and implemented control-flow-graph optimization passes to improve BPF execution speed.
• Extended tcpdump's transport-layer capture to work over IPv6.
• Contributed back to the open-source community — enhanced the "Caper expansion" section on tcpdump.org.
• Wrote a Go web server that visualizes step-by-step packet processing for user-uploaded pcap files (~2.1K requests/month).
• Received Honorable Mention from the Computing Research Association (CRA) for Outstanding Undergraduate Research.

GitLab — caper · bpfsimulator.com · Live on tcpdump.org

• Implemented the Raft consensus algorithm in Go using Test-Driven Development for robust leader election and log replication.
• Wrote automated tests that simulate network partitions and recoveries to verify fault tolerance across cluster nodes.
• Applied machine-learning techniques to tune election timeouts and heartbeat intervals, reducing leader re-election times.

Coursework spanning Data Structures, Algorithms, Systems Programming, Databases, Functional Programming, Computer Networking, Software Defined Networking, and Operating Systems.