Sovereign On-Premise Agentic AI Workbench using Open-Weight Multimodal LLMs for Confidential Industrial Work

Source summary

The proposed industrial AI workbench must run air-gapped on an organization's own GPU hardware, route at least two task types across replaceable open-weight models and keep all data local. It should perform multi-step work with sandboxed code, files, spreadsheets, local retrieval and multimodal OCR/vision, producing actual documents and other deliverables. A single-workstation demonstration may use smaller models when hardware is limited. Locality and model selection must be demonstrated, not assumed from self-hosting alone.