Kit Kyo · 乔磊

AI · Agent · LLM Infra Engineer ｜ Full-Stack

Profile

  • Kit Kyo · M · 28
  • Tokyo, Japan

Contact

Summary

AI Agent / LLM infrastructure engineer with a full-stack background. Focused on sovereign / on-prem AI agent infrastructure and edge LLM / MoE inference (deployability × data residency × close-to-metal). Experienced in building and operating multi-node k3s clusters from scratch, A2A message buses, unified LLM inference gateways, and observability stacks. 5 years of Java/Go full-stack foundation, now deep in LLM agent engineering and GPU inference optimization.

Education

  • Inner Mongolia Agricultural University — Computer Network Technology

AI / Agent / LLM

  • Claude Agent SDK
  • MCP
  • LangGraph / A2A
  • vLLM
  • Ollama
  • llama.cpp
  • MoE edge inference
  • Gemma / Qwen
  • RAG

Infrastructure

  • Kubernetes / k3s
  • NATS / JetStream
  • Docker / Kaniko
  • Tailscale / Liqo
  • nftables / PPPoE
  • Infisical
  • systemd / launchd
  • Linux (Debian/macOS)
  • vast.ai GPU

Backend & Full-Stack

  • Go (Gin/GoZero)
  • Java (Spring Boot/Cloud)
  • Python
  • TypeScript / Node
  • React
  • Vue
  • Three.js / WebGL
  • Flutter
  • MySQL / Redis
  • RabbitMQ

Experience

  • TLZS Corporation

    • Platform / Infrastructure Engineer (SRE)

  • Z-SOFT Corporation

    • Systems Development Engineer

  • Huanzhiyu Technology (Inner Mongolia)

    • Java Developer

  • Xunya Shikong Technology (Tianjin)

    • Java Developer (Intern)

Projects

  • Sovereign AI Cluster / Fleet Platform Engineering (SRE)

    • Stack:k3sPrometheus/GrafanalitellmKEDATailscale/Liqovast.ai A100
    • Built and operated a production-grade 8-node bare-metal k3s cluster from scratch (multi control-plane + etcd, 58 namespaces / 197 deployments); LVM storage, nftables/PPPoE networking, Tailscale + Liqo cross-cluster federation; 70 days of stable uptime, handling 60+ node/disk/network/GPU operations incidents.
    • Built a self-hosted observability + AI infrastructure stack: Prometheus/Grafana/Loki/Jaeger + Tetragon eBPF; litellm unified LLM gateway, Langfuse observability, KEDA autoscaling, Kaniko in-cluster CI; vast.ai A100 GPU orchestration (vLLM/Ollama, QLoRA fine-tuning of Gemma4); delivered business systems including jp-schedule-hub and jp-esign-forge.

  • Mamelog — Social Content Platform for Residents in Japan

    • Stack:GolangGoZeroFlutterKubernetes
    • RED/Xiaohongshu-style content-sharing app: distributed video upload/processing, marketplace and payments, AI content recommendation and Q&A powered by Ollama + Gemini; K8s circuit breaking/load balancing, supporting 100 concurrent users.

  • Integrated AI Tools Platform

    • Stack:GolangNextJSStable DiffusionChatGPT
    • Unified platform integrating multiple AI services such as image generation and conversational models; user/token balance management, remote multi-instance Stable Diffusion monitoring, ChatGPT API token distribution, supporting 100 concurrent users.

  • Government Grid-Management System

    • Stack:JavaSpring BootReactCesium
    • Community housing/population/grid-officer workflow management system, integrating GIS maps and a data dashboard; the front end uses Cesium for 3D building visualization that reconstructs real-world cityscapes.