Answers that trace back to a source.
Grounded LangGraph agents in a Neo4j knowledge graph so a fintech could trust them.
AI Engineer Consultant · core contributor on a fintech AI/Data team · Jun – Oct 2025 · remote, Dubai
Context
A fintech's AI/Data team needed bespoke GenAI and RAG for internal teams and external clients, built with product and data managers.
Problem
LLMs alone hallucinate and don't know company facts, reporting was manual, and inference spend had to stay under control.
Architecture
- Company documents & data
- LangGraph agents on FastAPI
- Neo4j knowledge graph · vector DBs
- Reports · newsletters · alerts
- HubSpot
Key decisions & trade-offs
Graph + vectors, not vectors alone
Integrated a Neo4j knowledge graph with vector databases to ground agents in company data.
Trade-offGraph modelling takes effort; in return answers are traceable, sustaining 90% accuracy.
Deliver into existing tools
Outputs went out as automated reports, newsletters and alerts, wired into HubSpot.
Trade-offLess control than a custom UI, but no adoption barrier: manual reporting time fell 60%.
Architect for cost
Advised on system architecture to balance accuracy against cost.
Trade-offNot every task gets the largest model; LLM/inference spend fell 30%.
Outcomes
- 90% answer accuracy sustained
- −60% manual reporting time
- −30% LLM/inference spend
- 5 proposed use cases taken into production