Aegis — Autonomous AI Incident Commander

Aegis is an autonomous AI Incident Commander designed to reduce Mean Time To Resolution (MTTR) during critical engineering outages. By integrating Agora WebRTC Conversational AI, Groq LPU LLM Acceleration, and Multi-Agent Telemetry Orchestration, Aegis converts complex incident triage into an automated, voice-guided command workflow. 

• Sub-600ms Diagnostic Latency: Groq openai/gpt-oss-20b engine processes diagnostic queries in ~560ms.
• Hands-Free Voice AI Channel: Talk directly to Aegis over WebRTC using Agora Conversational AI v2, featuring persistent audio streams across dashboard navigation.
• Parallel Sub-Agent Dispatches: Autonomous Logs, Metrics, and Topology sub-agents aggregate APM telemetry to pinpoint root causes with up to 98.4% confidence.
• Human-in-the-Loop Confirmation Gate: High-risk actions (Kubernetes auto-scaling, PagerDuty pages, Jira creation, Statuspage publishing) require explicit verbal or one-click IC authorization.

• Direct RTC Channel: Aegis joins the incident WebRTC channel as an active audio publisher and subscriber.
• Turn Detection Tuning: Configured with max_wait_ms: 2000, silence_duration_ms: 400, and interrupt_duration_ms: 160 to support natural speech interrupts.
• Global State Elevation: WebRTC audio tracks remain active at the root layout level (App.jsx), allowing engineers to switch between dashboard tabs while retaining continuous voice communication.

• Operates on Groq's dedicated LPU hardware (openai/gpt-oss-20b / llama-3.1-8b-instant).
• Context payloads are trimmed to 4 core key-value fields (Incident, Fact, Cause, Fix), eliminating prompt processing overhead.
• Features max_tokens: 150, warm Connection: keep-alive sockets, and an AbortController 800ms protection guard.

• Ingests telemetry logs, APM metrics, and support ticket spikes.
• Dispatches parallel sub-agents (Logs Agent, Metrics Agent, Topology Agent) to correlate failure cascades and calculate root-cause hypotheses with confidence metrics.

• Prevents unverified automated actions. Operations requiring authorization:
  • Kubernetes API: Scaling deployment replicas or rolling back containers.
  • PagerDuty: Escalating pages to secondary on-call engineers.
  • Jira Software: Generating post-incident tracking tickets.
  • Statuspage: Publishing customer-facing status updates.