Skip to content
MH
All projects
Data EngineeringGenerative AIAutomation

Call Processing Pipeline with Neo4j Graph AI

AI / Graph Databases / Data Engineering

An end-to-end AI-powered call intelligence system built in n8n that turns raw call webhook data into structured, actionable insight. It receives call events, extracts caller, callee, duration and recording metadata, and stores them in a Neo4j graph database.

Year
2025
Role
Automation & data engineer
Source
Public repository
Call Processing Pipeline with Neo4j Graph AI — concept diagram

Technologies

  • n8n
  • Neo4j
  • Cypher
  • LLMs
  • Webhooks
  • Data Pipelines

The problem

Call platforms emit events, not understanding. Who spoke to whom, how often, and what patterns exist across a whole account are relationship questions — and relationship questions are exactly what a row-per-call table answers badly.

The solution

Model calls as a graph. People, numbers and calls become nodes and edges in Neo4j, so "everyone this contact reached last month" is a traversal rather than a chain of joins. AI then reads the structured result and writes the insight layer on top.

Architecture

How it is put together.

01

Webhook ingest

Call events arrive at an n8n webhook as they happen.

02

Metadata extraction

Caller, callee, duration and recording metadata are pulled out and normalised.

03

Graph modelling

Entities and their relationships are written into Neo4j as nodes and edges.

04

Graph queries

Cypher queries surface connection patterns, frequency and reach across the graph.

05

AI insight generation

An LLM reads the structured query results and produces readable, actionable insights.

Key features

What it actually does.

  • Real-time webhook ingestion of call events
  • Normalised metadata extraction
  • Neo4j graph model of people, numbers and calls
  • Cypher-based relationship analysis
  • AI insight layer grounded in structured results

Engineering challenges

The parts that were hard.

Identity resolution

The same person appears as several numbers and formats. Normalising before writing to the graph prevents one contact from becoming five nodes.

Grounding the AI layer

Insights are generated from query results rather than raw events, so the model summarises facts the graph has already established.

Outcomes

What came out of it.

Qualitative outcomes only — no invented benchmarks or metrics.

01

Call history queryable as relationships instead of flat rows

02

Insight generation grounded in structured graph output

03

A pipeline that runs continuously from webhook to insight

Next project

End-to-End Sales Forecasting Data Warehouse

Python cleaning, a SQL Server warehouse, SSIS ETL, ML forecasts and Power BI dashboards.

Let’s build

Want something like this for your business?

Tell me the problem and I will come back with an architecture, a scope and a timeline.