In small and medium-sized manufacturing enterprises, large volumes of heterogeneous data are generated on a daily basis—ranging from requirements and design documents to machine and sensor data. The use of conventional cloud-based large language models (LLMs) to analyze this data often fails due to data confidentiality concerns as well as high hardware and energy costs.
The project demonstrates that current reduction techniques can compress LLMs to ≤ 10 billion parameters and enable their execution on workstations with 4–8 GB VRAM through four-bit quantization. Compared to conventional 16-bit models, this reduces hardware and energy requirements by more than 70 %, while ensuring that all data remains within the company network.
Based on this technical foundation, two prototype assistance systems are developed. A requirements co-pilot automatically checks specifications and requirement documents for inconsistencies, references to standards, and gaps in prioritization. A machine condition co-pilot integrates log, image, and sensor data, derives hypotheses about root causes, and generates recommended actions.
The project envisions the development of an end-to-end process chain—from data selection through compression, quantization, and graph-based retrieval to deployment monitoring. This will provide a transferable, resource-efficient, and data-sovereign AI foundation, enabling especially small and medium-sized industrial enterprises to adopt practical language model solutions.
