Conversational ITSM Intelligence Agent

A multi-agent conversational platform that answers operational questions in plain English and automates root-cause analysis across incidents, changes, and problems.

All Systems

The Problem

Engineering and support teams needed natural-language access to incident, change, and problem data scattered across a live ITSM system and its historical data warehouse, plus manual, time-consuming root-cause analysis linking failed changes to the incidents they caused.

Why Traditional Systems Failed

Generic chatbots lack ITSM domain context and the guardrails an enterprise support org needs. Naive LLM-to-SQL tools are unreliable for production data because they let the model both query and interpret data in one uncontrolled step, risking hallucinated numbers in executive-facing reports.

Multi-Agent Architecture

Built a multi-agent pipeline on Google's Agent Development Kit: a Router agent classifies intent, a deterministic Prompt Expander normalizes the request, an Understanding agent extracts structured parameters, a Generation agent drafts SQL or live API calls, a Review agent validates them, and only then does a pure-Python Execution stage run the query — with zero LLM involvement in that step, so data never round-trips through the model. A separate Reasoning agent turns validated results into narrative summaries, tables, and charts. Dedicated two-agent pipelines handle live incident and change-request lookups against the source ITSM system, with a self-correcting retry loop (up to three attempts) if generated SQL fails.

Interactive Architecture

Hover a node for detail · click to isolate its connections

Access & Security

Agent Pipeline

Data & Knowledge

Ingestion & Sync

LLMs handle understanding and narration only. SQL execution and RCA logic run in deterministic Python. BigQuery remains the system of record.

▸ TRY IT LIVEReplay this pipeline in the Lab — pick a question, watch the retry loop save the run

Key Capabilities

Secure Natural Language Querying

Lets users ask operational questions in plain English across incidents, changes, and problems.

Multi-Agent Reasoning Pipeline

Router, Understanding, Generation, and Review agents collaborate before any query touches production data.

Deterministic Data Execution

SQL and live API calls run in pure Python — the LLM never sees or touches raw data, only validated results.

Automated Root-Cause Analysis

5 Whys synthesis and change-caused-incident (CCI) detection automatically link failed changes to the incidents they caused.

Scheduled Enterprise Sync

Incident, change, and problem data syncs from the source ITSM system on a fixed schedule to keep the warehouse current.

What This Enables

  • Faster incident context retrieval during outages
  • Automated root-cause narratives instead of manual ticket archaeology
  • Self-service operational insight for non-SQL stakeholders
  • Executive-ready summaries generated directly from validated data
  • Safer LLM adoption pattern: reasoning and narration only, never raw data handling

What This Is Not

  • Not a generic AI chatbot
  • Not an autonomous remediation system
  • Not a replacement for the core ITSM system of record
  • Not a system where the LLM ever executes queries against production data directly

Business Value

Democratized access to ITSM data for non-SQL users, automated root-cause synthesis that used to require manually cross-referencing incidents against recent changes, and gave leadership executive-ready narrative summaries instead of raw ticket exports.

Key Learnings

  • Separating deterministic code from probabilistic LLM steps is the key reliability lever: schema injection, date parsing, and data execution stay in plain Python so numbers in reports are never model-generated.
  • A self-correction loop on generated SQL (validate, retry, cap at a few attempts) meaningfully improves first-try accuracy without needing a human in the loop.
  • Root-cause analysis features (5 Whys, causal linking between changes and incidents) deliver more perceived value than the query interface itself — teams cared more about "why did this happen" than "show me the data."