AI Agents · LLM Systems
AI Ops Agent
An AI agent system in Microsoft Teams automating incident response, postmortems, analytics and operational workflows across multi-cloud tenants.
4
specialist agents
22
tools
Multi-tenant
isolated partner chats
Overview
An internal assistant that lets engineers file and track incidents, write postmortems, look up policies and get answers from operational data - all from a chat, without forms. Partners use the same bot in their own chats.
An LLM intent router dispatches each message to a specialist: a slot-filling agent for multi-turn intake, function-calling agents that each encapsulate their own tools, a RAG pipeline for policies and a guarded natural-language query engine.
Highlights
- Intent routing to specialist agents; each encapsulates its own tools and prompt, so no model call sees another domain's tools.
- SLA monitoring and escalation on scheduled Azure Functions.
- Grounded policy answers: parent–child chunk retrieval over pgvector, citations from retrieved sections only, refusal below a similarity threshold.
- Data questions in plain language, with defense in depth: every generated query is validated, run read-only and time-limited.
- Tenant isolation for partner chats: scope resolved from the conversation and enforced before routing.
Architecture
Pick a scenario to trace a message through the system
billing server is down, keeps returning 502
Triggers
Handlers
Data & integrations
billing server is down, keeps returning 502
Inference pipeline
Where the model runs, and what constrains it
- 01
Input guardrails
- tenant scope resolved from the conversation
- commands and active dialogs short-circuit routing
- 02
Router
- intent classification
- schema-constrained JSON output
- 03
Context construction
- RAG: parent–child chunks in pgvector
- data questions: schema, glossary, previous question
- dialog state and people directory
- 04
Generation
- function calling over a scoped tool set
- slot filling with per-field confidence
- 05
Output guardrails
- citations only from retrieved sections
- refusal below a similarity threshold
- generated queries validated, read-only, time-limited
- field allowlists for partners
- 06
Actions
- incident writes and Planner sync
- timesheet and leave requests
- PDF / Excel behind expiring links
Aruarian Dance
Nujabes