Senior Software Engineer, Inference Engine (Platform Software)

FuriosaAI · Software · Seoul, South Korea
공고 번호 4005788201

LLM 추론 엔진의 batching·KV cache·분산 실행과 처리 성능 개선.

학사 명시 · 도전 지원

확인할 점: 3년 관련 경력 또는 동등 역량·Rust/C++ 필수. 엔진 내부/KV-cache/분산 추론 구현 미확인.

영문 이력서 PDF ↗영문 이력서 Overleaf ZIP ↓

Cover Letter PDF ↗Cover Letter Overleaf ZIP ↓

이 공고의 지원서 열기 ↗

Career Summary

필수 Career Summary 칸에 붙여 넣기.

I am an ML engineer and AI Researcher at ISPARK R&D Center responsible for TieMing's core AI features, from model training to inference APIs and deployment. I bring a multimodal research background—five publications, four as first author—and turn experimental results into working products. I operate vLLM services and improve how GPU workloads progress, use resources, and deliver results.

I build and operate vLLM and FastAPI model services with Docker deployment and service/GPU monitoring. I ran distributed training on two RTX 4090 GPUs with checkpoint recovery and now also use two A100 GPUs through NIPA's support program.

I reduced server processing time by 36.9% on a 30-minute recording through speech-region batching. I also redesigned long-audio processing as recoverable asynchronous jobs, separating completion and recovery from the HTTP request lifetime.

I bring practical experience operating model services, adapting task-specific models, and investigating GPU processing behavior. Working across these layers lets me connect a serving symptom to the workload and execution conditions beneath it. At FuriosaAI, I can contribute workload analysis and model validation to inference engine development. I would trace a request through execution, reproduce the limiting condition, and use reference outputs and timing measurements to assess implementation changes.

Cover Letter

PDF 첨부 또는 Enter manually에 붙여 넣기.

Dear FuriosaAI Hiring Team,

I am applying for the Senior Software Engineer, Inference Engine (Platform Software) position (Job ID 4005788201). I bring practical experience operating model services, adapting task-specific models, and investigating GPU processing behavior. Working across these layers lets me connect a serving symptom to the workload and execution conditions beneath it. Before ISPARK, I served one year as an information-systems NCO in the Republic of Korea Air Force (Dec 2021–Dec 2022), maintaining Linux servers through shell scripting and log analysis.

I take responsibility for model deployment as well as training, building vLLM and FastAPI inference services in Docker with Prometheus and Grafana monitoring. Watching service failures alongside GPU usage helps me distinguish application problems from resource constraints. I ran distributed training on two RTX 4090 GPUs with BF16 precision and checkpoint recovery, and now also use two A100 GPUs through a national computing-support program. I track experiments in Weights & Biases so model changes can be compared across runs.

I redesigned long-audio processing and reduced server processing time by 36.9%, from 122.6 to 77.4 seconds on a 30-minute recording, using Silero voice activity detection and faster-whisper batching at batch size 8. The same workload had also exposed a delivery problem: the server finished after the synchronous HTTP request timed out. I separated live sessions from recoverable asynchronous jobs, with task polling and persisted progress. I treated recognition quality, execution time, and delivery as separate engineering problems, allowing each change to address a specific cause.

At FuriosaAI, I can contribute workload analysis and model validation to inference engine development. I would trace a request through execution, reproduce the limiting condition, and use reference outputs and timing measurements to assess implementation changes.

Sincerely,
Jeonghun Kim