12 hours ago
Seoul, Korea, SouthSenior
Responsibilities
- Build and operate a serving platform and control plane on Kubernetes-based GPU clusters.
- Integrate and improve inference engines and orchestration layers, including request routing, autoscaling, and multi-model operation.
- Create safe model onboarding, performance and quality regression testing, deployment automation, canary release, rollback, authentication, quota, and rate-limiting capabilities.
- Establish and operate SLO-based observability for latency, throughput, availability, and error rates, and participate in incident response and postmortems.
- Manage GPU allocation, isolation, and prioritization policies and measure GPU utilization and per-request or per-token costs.
- Improve routing, scheduling, caching, and model integration logic and validate serving-engine features under production traffic.
- Validate and productionize optimization techniques such as quantized checkpoints and speculative decoding.
Requirements
- At least 5 years of engineering experience, including at least 3 years in infrastructure, SRE, platform engineering, or ML systems.
- Direct experience operating production services where incidents affect users.
- Experience designing and operating deployment, load balancing, scheduling, and autoscaling in Kubernetes or an equivalent orchestration environment.
- Experience building metrics-, logs-, and traces-based observability and structurally improving root causes through on-call response and postmortems.
- Experience implementing operational automation or platform functions in code and managing infrastructure and deployments reproducibly with IaC, GitOps, and CI/CD.
- Experience aligning requirements with researchers and product developers and documenting design and operational decisions.
- No restriction preventing overseas business travel.
- Experience with large-scale or low-latency real-time services, capacity planning, quotas, prioritization, or resource isolation is preferred.
- Experience serving LLM or ML models in production on GPUs using vLLM, SGLang, TensorRT-LLM, or Triton Inference Server is preferred.
- Understanding and practical application of scheduling, batching, KV cache management, continuous batching, prefill-decode separation, prefix caching, or KV cache offloading is preferred.
- Experience analyzing and improving multi-node GPU or system-level performance bottlenecks is preferred.
Benefits
- Permanent employment is offered, although the employment type may change based on the negotiated terms.
- The workplace is Yeoksam Centerfield.
- A 5-month probationary period applies, with no changes to employment type or salary during the period.
- The hiring process includes document screening, a phone interview, possible assignment, personality assessment, technical-fit interview, and two to three culture-fit interviews; steps may vary by position or candidate.
- Hiring is ongoing and may close early when the position is filled.
- Applicants eligible for legally recognized employment-protection programs, including people with disabilities and distinguished service to the state, receive preference under applicable law.
Tech Stack
Categories
About KRAFTON
KRAFTON is a Seoul-based, KRX-listed video game company that develops and publishes PC, console, and mobile titles. Its studios include PUBG Studios, Bluehole Studio, Striking Distance Studios, and Unknown Worlds, best known for the PUBG: Battlegrounds franchise (and PUBG Mobile with Tencent) and games like Subnautica and TERA. The company monetizes through premium sales and live-service content across global markets.