About

Jihwan Park

I studied computer science and electronic engineering at Kyung Hee University, and I organize my work around reliable infrastructure for constrained environments. The recurring question is how services and AI workloads behave when network access, storage recovery, GPU placement, memory, or hardware resources are limited.

The public version of this notebook focuses on constraints, architecture choices, recovery paths, and implementation details while omitting private infrastructure details.

Restricted Infrastructure

Selected Docker-based services were moved toward a K3S-based internal platform, with private image distribution, service exposure, persistent storage, and deployment repeatability as the main concerns.

Recovery and Reliability

Stateful services made recovery boundaries visible: PVC detach/attach behavior, rebuild time, readiness, operator action, and realistic RTO/RPO expectations.

Resource-Constrained AI Systems

Graduation projects looked at AI workloads from the systems side: TFLM tensor memory reuse on constrained devices and bit-shift CNN acceleration on Zynq FPGA with area/runtime tradeoffs.

Focus

What I Actually Track

The site preserves the concrete constraints, failure modes, and design decisions behind each note.

Dependency Paths

Which images, artifacts, packages, and bootstrap services must be available when external access or the primary registry is unavailable.

Recovery Timelines

What has to happen between a node or storage failure and service readiness: pod scheduling, volume movement, replica rebuild, health checks, and operator action.

Execution Constraints

How AI workloads change when the limiting factor is not only model accuracy but memory arena size, FPGA area, GPU allocation, dataflow, or degraded serving.

Boundaries

Scope and Boundaries

The goal is to stay specific about engineering scope without overstating infrastructure, recovery, or AI claims.

Beyond A Kubernetes Installation

The point is not simply that Kubernetes was installed. The useful part is understanding how restricted-network service operation changes when image supply, storage, service exposure, and rollback paths must be designed together.

Recovery, Not Zero-Downtime Claims

Longhorn-backed PVCs do not automatically imply zero downtime. I frame recovery in terms of detach/attach behavior, rebuild time, readiness, operator action, and RTO/RPO boundaries.

Systems-Oriented AI Work

The AI thread is operational and systems-oriented: full-GPU serving constraints, degraded availability, TinyML memory planning, and hardware-aware acceleration.

Sanitized Public Notes

Operational details are generalized. The public artifact should show the engineering problem and tradeoff without exposing private hostnames, internal IPs, credentials, or raw logs.