Multi-Agent Digital Assistant for Marketing
Extendable digital analyst - answers natural-language questions over a marketing graph with charts, profiles, and evidence-backed recommendations.
Why it mattersA digital analyst that joins ambiguous business questions to schema-safe Cypher queries, charts, and human-readable narrative - so growth teams act on data instead of waiting on a data-science queue. The same supervisor + specialist-team pattern is portable to any domain with rich relational data.
What it does
A supervisor-orchestrated multi-agent system that reasons over a marketing graph database to answer natural-language questions, generate charts, profile audiences, and return evidence-backed recommendations. Ask "which audience segments drove conversions in the last campaign?" and the system routes the question through the right specialist team, queries the graph, renders a chart, and stitches the result into a readable narrative.
Where it applies
- Growth teams running attribution, segmentation, or paid acquisition at scale who need an analyst on demand, not a data-science ticket queue.
- Marketing-ops groups whose insights get lost in translation between data warehouses, audience graphs, and campaign platforms.
- Any organisation with rich relationship data that struggles to act on it because the path from question to chart is too slow.
How it works (high level)
A supervisor agent (LangGraph) classifies each incoming question and routes to one or more specialist teams: a retrieval team that translates the question into schema-aware Cypher and runs it against Neo4j, a profiling team that builds user-level summaries via multi-hop graph traversals, a visualisation team that writes and executes Python to render charts, and an interpretation agent that turns the graph result into readable narrative with citations back to the underlying data. Every agent call, tool use, and supervisor decision is traced through LangFuse so the reasoning is auditable end-to-end.
Why this design
Multi-agent systems look tidy in diagrams and get messy in production. The architecture is also a study of three failure modes I wanted to characterise: supervisor misrouting, schema drift in generated Cypher, and how end-to-end latency decomposes across agents under load. The repo doubles as an empirical platform for those questions, not just a working tool.
Outcome
A natural-language analyst that turns marketing-graph data into evidence-backed insights and concrete next steps, with the supervisor pattern intentionally extendable so new specialist teams (forecasting, anomaly detection, creative generation) can be plugged in without rewriting the core.
Stack
Python · LangGraph · Neo4j · LangFuse · Docker Compose · Streamlit · matplotlib.