{"slug":"german-engineering-consultancy-runs-a-self-hosted-qwen3-ai-platform-on-four-nvidia-l40s-gpus-for-200-employees","url":"https://findausecase.com/use-cases/german-engineering-consultancy-runs-a-self-hosted-qwen3-ai-platform-on-four-nvidia-l40s-gpus-for-200-employees","title":"German engineering consultancy runs a self-hosted Qwen3 AI platform on four NVIDIA L40S GPUs for 200 employees","description":"Particula Tech built a fully self-hosted AI platform for a 200-employee German structural engineering consultancy: Qwen3 models served with vLLM on a single server with four NVIDIA L40S GPUs, Open WebUI with Keycloak SSO, local RAG over project archives, Whisper transcription and a code assistant. Motivated by client NDAs and German data protection rules; the source reports replacing EUR 14,000/month of cloud AI spend with EUR 48,000 of hardware.","company":"German engineering consultancy (unnamed client of Particula Tech)","industry":"Professional Services","country":"Germany","aiCapabilities":["Generative AI","Conversational AI"],"technology":["Qwen3","Qwen2.5-Coder","vLLM","Open WebUI","Whisper","Qdrant","Ollama","Keycloak","NVIDIA L40S"],"deployment":"On-Premise","problemStatement":"About 80 employees had personal ChatGPT accounts they were pasting client data into, while monthly cloud AI spending had reached EUR 14,000 with no central oversight of what data was leaving the network. Client NDAs on most infrastructure projects prohibited sending project data to third-party cloud services, and German data protection regulations added a further constraint.","solutionApproach":"Particula Tech built a self-hosted platform on a single server with four NVIDIA L40S GPUs running open-source Qwen3 models served with vLLM, with Open WebUI as the company-wide interface behind Keycloak SSO, a local RAG knowledge base in Qdrant over the firm's project archives and the DIN and Eurocode standards, local Whisper transcription with meeting summaries, and a Qwen2.5-Coder-32B code assistant.","businessValue":"The source reports the platform eliminated EUR 14,000 per month in cloud AI subscriptions and API costs and, with EUR 48,000 of hardware and about EUR 400/month in electricity and maintenance, paid for itself in 3.5 months. It reports adoption of 94% of employees by month three. The firm's estimate of about 15 hours saved per engineer per month on research is described by the source as self-reported and likely optimistic.","evidence":{"band":"high"},"sourceUrl":"https://particula.tech/portfolio/local-ai-platform-engineering","dates":{"publishedAt":"2026-10-26T05:46:12.676Z","publishedAtSource":"pipeline","updatedAt":"2026-10-26T05:46:12.676Z"},"license":"Open for reading and citing with a link to https://findausecase.com/use-cases/german-engineering-consultancy-runs-a-self-hosted-qwen3-ai-platform-on-four-nvidia-l40s-gpus-for-200-employees. Bulk republication requires permission."}