AI최고급신진연구자지원 (AI스타펠로우십, 지역주도형)

진단 AI 에이전트의 경량화 및 추론 최적화를 담당하는 과제이다. 양자화, 프루닝·지식 증류, 컴파일러/커널 최적화, 서빙·런타임 최적화를 통해 대규모 멀티모달 에이전트를 실용적인 지연시간과 메모리 예산 내에서 서비스하는 것을 목표로 한다.

  • This project is funded by IITP (Institute of Information & Communications Technology Planning & Evaluation), Korea, under the AI Star Fellowship program.
  • Period: Jul. 1, 2026 – Dec. 31, 2031 (6 years).

Within this project, our lab is responsible for the lightweighting and optimization of the diagnostic AI agent.

Diagnostic agents combine multimodal perception, clinical language-model reasoning, and tool orchestration, which makes them large and latency-sensitive — precisely the regime where naive deployment becomes impractical. Our role is to close that gap:

  • Quantization — low-precision weights and activations that preserve diagnostic accuracy.
  • Pruning and distillation — smaller models at comparable quality.
  • Compiler and kernel optimization — hardware-aware operator tuning for the target accelerators.
  • Serving and runtime optimization — meeting throughput, memory, and latency budgets in real deployments.

The result is a diagnostic agent that can be served on-device or at the edge, so that inference stays fast and affordable at the point of care.