Shanghai Industry's "New Journey": ManuDrive Industrial LLM Enters the Workshop — Industrial Time-Series Control Model Powers New Industrialisation

Title

Shanghai Industry's "New Journey": ManuDrive Industrial LLM Enters the Workshop — Industrial Time-Series Control Model Powers New Industrialisation

Content

Chuanning Bio, a top Chinese maker of erythromycin thiocyanate, cephalosporin, and penicillin intermediates that together dominate the relevant domestic sub-markets, ran 500-tonne fermentation tanks where one batch could only convert less than 1% of input into final product. The fermentation cycle lasted 1-4 weeks in a strongly acidic environment, so the plant had to staff 24-hour shifts to monitor each tank. Despite holding the top global position in its core fermentation products, the company struggled with the variability and human-dependency of the process. Professor Li Jinjin, director of the Artificial Intelligence and Microstructure Lab at Shanghai Jiao Tong University, embedded her team on site in Xinjiang for 9 months, working at desks set up next to the four-storey fermentation tanks in 40°C heat. They built ManuDrive, an industrial time-series control LLM that predicts the next 180 hours of fermentation parameter curves with 99.9% accuracy and runs a predict-optimise-execute-feedback closed loop on top of classical control theory. The model is paired with a Frontline Deployment Engineer (FDE) delivery model: instead of writing code in an air-conditioned office, the engineering team is stationed at the customer site to embed the model in the real production workflow. After ManuDrive took over the 500-tonne Chuanning Bio fermentation line, output rose 3-5 percentage points, production volatility dropped 50%, and the single line added at least RMB 30M per year. Compared with a classic AI deployment that needs massive data and GPU clusters, ManuDrive runs on a dozen commodity GPUs using only 5% of the data, which makes it accessible to small and mid-sized manufacturers. The system is built as 127 modular "industrial Lego" components (perception, planning, memory, tool use, etc.) that can be reassembled for new scenarios. An algorithm originally developed for microbial metabolism in fermentation was packaged as a module and reused in wastewater treatment, lifting cold-start efficiency by 90% and reducing per-tonne treatment cost by another 10%. The same playbook has expanded to CAD drawing generation (cutting single-drawing time to under 3 seconds for Tai Heavy Industry Group, with net profit up more than 17%), industrial building design (Shuangliang Group cut design verification time from weeks to 10 minutes, with net profit up more than 15%), textile plate-making, scheduling, and waste incineration. If the system reaches 30% of China's core fermentation capacity, the new industry-wide value could exceed RMB 15B per year.