Agentic Vehicle Intelligence
Specialised agents search the web for vehicle prices, validate them, normalise currencies and persist the result — with provenance — in a Neo4j knowledge graph.
Autonomous research becomes valuable when its output is structured, traceable and reusable.
01Context
Registration datasets tell you what was sold, not what it costs. Market prices for niche European vehicles are scattered across manufacturer, dealer and marketplace sites.
02Problem
Enrich thousands of vehicle records with a defensible price — not a single hallucinated number from one LLM call.
03Constraints
- Free web sources only
- Prices appear in many currencies and formats
- Sources vary widely in reliability
- Must run unattended in batches
04My role
What I personally built
- Agent hierarchy and responsibilities
- Search, scraping and price-validation logic
- Confidence scoring by source type
- Currency normalisation to EUR (original currency preserved)
- Neo4j schema and graph enrichment
- Batch processing and progress tracking
What others owned
Independent project — no team. Everything here is mine.
05Architecture
Architecture explorer
The business flow — what happens, in plain words.
01 / 05
Vehicle list — Models that need a market price.
06Key decisions
Specialised agents instead of one research prompt
- Because
- Each step can be tested, logged and replaced independently.
- Trade-off
- More orchestration code and state to manage.
Persist to a knowledge graph, not a CSV
- Because
- Price ↔ source ↔ vehicle provenance becomes queryable.
- Trade-off
- Requires running and modelling Neo4j.
07System
A coordinator fans out query variations to a search agent, a scraper extracts candidate prices, a validator scores them by source type, a currency agent converts to EUR, and a graph agent writes Vehicle, Price and PriceSource nodes with their relationships.
08Challenges
- Separating list prices from monthly offers and accessories
- Scoring confidence without a ground-truth dataset
- Keeping the original currency for auditability
09Outcome
A working R&D pipeline that enriches vehicle records in configurable batches and stores every price with its source and confidence in a queryable graph.
10What I learned
Decomposing an agent into narrow responsibilities makes each failure visible — and fixable — instead of hidden inside one prompt.
11Stack
- Python
- Multi-agent
- DuckDuckGo search
- Web scraping
- Neo4j
- Cypher
- Docker
- YAML config
- pytest
12Evidence
- Private repository — walkthrough on request
- Neo4j — Agentic Knowledge Graph Construction (certification)