Zero-GPU Autopilot: Orchestrating Kagent and Kgateway for Private, Se...
summary
Claims a private, CPU-only autopilot that uses vLLM + Kagent with OpenVINO/IPX to reason, and Kgateway policy-enforced reroute to enforce safe actions. Demo injects faults into a canary, fires a Prometheus alert at ~15% 5xx under 3,000 requests, and the system patches the HTTPRoute weight from 80→0 weight change (80/20 → 100/0) to stop failures. Architecture is air‑gapped, SOP-driven, closed‑loop automation that runs without GPUs, applies declarative patches, and visibly reduces error rate and MTTR in the demo.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 2:07Kubernetes troubleshooting problems
- 4:13Private zero-GPU autopilot architecture
- 8:52VLLM and K agent for local reasoning
- 13:25K gateway policy enforcement
- 16:35Zero-GPU cost and scalability
- 17:36Live demo setup
- 21:04Fault injection and alerts
- 22:08Automated debugging workflow
- 25:17Handling hallucinations
- 28:03Cost comparison with GPUs
- 30:00Q&A
related talks
42mLLMs on Autopilot: Running AI Agents on Kubernetes With Open Source Tools
31mBeyond the Copilot: A Blueprint for Autonomous Mobile Reliability
26mAI Traffic Is Different: Rethinking Proxy Arc...
30mLLMs on Autopilot: Running AI Agents on Kubernetes with Open Source Tools
35mMulti-Layered Guardrails for Cloud Native AI: Enforcin...
34mOpen Source AI Agents on User-Owned Infra: K3s, MCP, & GPU Sharing in Practice
01h 42mBuilding Responsible AI Agents in Decentralized Open-Source Environments
30mWhy Prompt-Level Guardrails Fail for AI Agents