AI agents & RAG applications

Agents and RAG systems are only as good as their data foundation and their evaluation. I design and build both together: data pipelines that reliably extract relevant content into a vector store, and agent logic that answers reliably and triggers actions on top of that foundation.

A typical project moves through several stages: connecting the data sources and building the retrieval pipeline, then prompt engineering and tooling for the agent logic, followed by integration into existing frontends or products. Evaluation runs throughout to confirm the answers are actually correct.

For Silver Investment Partners, I built a RAG-based pipeline that automatically processes information memoranda and uses Qdrant as the vector store to extract and structure comparable company metrics. For COMAN Software GmbH, I integrated a Q&A chatbot built on local LLMs (Ollama, LM Studio, NVIDIA DGX Spark) with Qdrant as the retrieval layer directly into existing software products. And for leezy.ai, I built the RAG pipelines on PostgreSQL with pgvector, combined them with LangChain.js and Hugging Face, and developed an agent that automatically detects whether a user wants to book an appointment or trigger another action.

The result is systems that hold up in production, not just in a demo.