Case studies
Systems I have built, and how.
Five production systems. Each one covers the problem, the architecture, the decisions behind it and what went wrong. Client names are withheld.
01AutomotiveLive
Dealer service agents
Multi-agent system and data platform
Dealer staff called or emailed HQ for specifications, warranty answers and order status, and waited. The system now carries around 200,000 questions a week, each one scoped to the dealership asking it.
Questions a week across 5,000 dealerships, each scoped to the dealer asking
- Azure OpenAI
- LangGraph
- pgvector
- Azure Functions
- Apache Superset
02RailLive
Requirement compliance platform
Agents for a safety-critical workflow
Requirement review took 30 to 50 days per document. The pipeline produces a first full analysis in under five minutes, and an engineer signs every compliance position.
To a first full compliance analysis, down from 30 to 50 days
- Azure AI Document Intelligence
- Azure OpenAI
- pgvector
- LangChain
- FastAPI
03PharmaLive
Data quality portal and analytics platform
Data quality gate and analytics platform
Errors from 26 markets were found downstream, in management reports and regulatory submissions. They are now caught at the landing zone, ticketed to the owning country, and data quality incidents fell by around 60 percent.
Fewer data quality incidents across 26 regulated markets
- Microsoft Fabric
- Power BI
- Azure Data Lake Storage
- SQL database
- Data quality portal
04ManufacturingLive
Formulation and offer assistant
Retrieval assistant with evaluation
A first reply to a customer took at least two weeks. It now comes in the same session.
First reply to a customer, down from at least two weeks
- Azure OpenAI
- pgvector
- FastAPI
- React
- Azure Container Apps
05FinanceLive
Peer and transaction recommendation agents
ML ranking, LLM re-ranking and an MCP server
Analysts assembled peer and precedent lists by hand across three databases, and two analysts rarely produced the same list. The agents now suggest ranked candidates with reasons; the analyst decides.
To a ranked peer shortlist, previously assembled by hand across three databases
- Python
- LangChain
- AWS Bedrock
- Snowflake
Contact
Discuss a project.
Tell me what you are building and what is in the way.