PlaceAI automated backoffice knowledge search using RAG+MCP and hybrid search.
By tuning YARN's RESERVED Resource and Trino settings, available memory was increased by about 40%.
Using digital topographic maps and a Street View SegFormer, narrow roads were identified more accurately.
By bundling per-Task failure emails in Airflow into a single message, alert noise was reduced.
Docker Compose was used to align the local Airflow environment more closely with the deployment environment, enabling debugging in PyCharm as well.
By separating tagging logic from storage and adding a write buffer, updates were reduced by 67%.
NAVER GLACE is upgrading its place and review services with 15+ AI models and PlaceLM.
Boosted CPU serving performance 4x through torchserve thread configuration and ipex pinning.
Used GitHub Actions and Marketplace actions to automate PR approvals, notifications, and syncing.
HA was verified with 4-node distributed MinIO, and archived data was operated with Backup & Recovery.
AI Briefing