The AI I care about is not a demo: it is a feature in production that reads and classifies documents, answers questions with its sources and acts inside the software the teams already use.
In a management software suite built on microservices, whose interfaces and document management I lead, AI is in production: OCR and automatic classification of incoming documents, with their data extracted; a RAG assistant that answers users’ questions from the software’s documents and content; self-hosted LLMs with vLLM; AI agents that chain actions on its APIs. I run the workshops with the client to align technology and business, one use case at a time.
The models are self-hosted and served by vLLM, so that clients’ documents stay on the infrastructure. RAG answers link back to their sources, and the agents only use the tools within their scope.
I start from the process: which document comes in, who reads it, what decision it triggers. Then I pick the simplest technique that holds: OCR and rules when they are enough, a language model when documents vary too much, RAG when the answer is in the documents, an agent when the task chains several actions.
For documents: OCR with Tesseract and OCRmyPDF, classification, barcode reading, matching each scanned document to its order and its document type, structured extraction. For RAG: document chunking, embeddings, vector search and generation with sources. For models: vLLM for self-hosted LLMs, Anthropic Claude, Mistral, Ollama, the OpenAI API and LangChain, TensorFlow for machine learning. For integration: Java / Spring Boot, Kafka and REST APIs, so that AI fits into reliable processes.
