From customer complaint to NOC ticket
in seconds.

Multi-agent assistant on Google ADK + Vertex AI Gemini. BigQuery scans partitioned outage history. AlloyDB handles CDR lookups via NL2SQL and persists the ticket.

4 ADK sub-agents · SequentialAgent root · MCP Toolbox · BigQuery · AlloyDB

Launch NetPulse AI View on GitHub
/ 01 — How it works

Four agents, one ticket.

A complaint enters the chat. Four ADK sub-agents run in sequence — each one reads the shared session state, queries its data store, and enriches the state for the next. The fourth writes the ticket and the run ends.

Full architecture
IN Customer complaint Flask · SSE chat
→ prompt customer_complaint
01 Classifier Tags category & region from raw complaint text.
Gemini native tool
→ emits category, region
02 Network Investigator Pulls related outages from BigQuery.
Gemini BigQuery · MCP
→ emits network_findings
03 CDR Analyzer Queries call records via natural language.
Gemini AlloyDB NL2SQL
→ emits cdr_findings
04 Response Formatter Synthesises & persists the incident ticket.
Gemini AlloyDB · INSERT
→ emits incident_tickets
OUT Incident ticket AlloyDB · NOC dispatch
→ persists incident_report
/ 02 — Data Viewer

The data behind every ticket.

Read-only views into the live sources the agents query — for poking at edge cases, validating fixtures, or seeing what the system saw before writing a ticket.

All sources
BigQuery · partitioned

Network Events

Outage and maintenance events across 10 cities, day-partitioned and clustered by region/severity.

50,000 rows AlloyDB · NL2SQL

Call Records

Call Detail Records with anchor-clustered drop and failure events around outage windows.

5,000 rows AlloyDB · write

Incident Tickets

The agents’ output destination — every saved ticket lives here with category, severity and recommended action.

live writes
/ 03 — Resources

Read the build.

Documentation, the full architecture deep-dive, and a guide to wiring NetPulse against your own telco data.

All resources
Reference

Documentation

What NetPulse is, what each agent does, and how the pieces fit together. Single-page, plain prose.

7 sections Deep-dive

Architecture

Subsystems, data flow, the model failover ladder, and the non-obvious choices that keep the system stable.

end-to-end flow How-to

Bring Your Own Data

Map your operator’s CDR + network events tables into the agent toolset. Schema, prompts, and the gotchas we hit.

schema + prompts