On-premise AI exploration for SMEs

Problem: an SME wanted to use AI for document analysis but refused to send confidential data to cloud APIs (ChatGPT, Claude). Solution: deployment of a Mistral 7B model on-premise on dedicated GPU (local server), with RAG interface to query the internal document base. Result: functional AI with no data leaving the IS, operating cost under 200 euros/month, performance suited to the use case. This case study details the context, architecture choices, delivery steps, observed outcomes and follow-up criteria agreed with the client. Powehi documente les décisions, les contraintes de sécurité, la méthode de déploiement, les indicateurs de réussite, les enseignements et les prochaines étapes afin que chaque exemple puisse être comparé à une situation réelle de PME.