After ManuDrive was connected to Chuanning Biotechnology's 500-tonne fermentation tank, output rose about 5%, production volatility fell by 50%, and a single 100-tonne-class line gained at least 30 million yuan a year in efficiency. The model was developed by Shanghai Jincheng Technology, a company founded by Professor Li Jinjin, director of the Artificial Intelligence and Microstructure Laboratory at Shanghai Jiao Tong University. The page calls ManuDrive the international first AI time-series control large model.
Li said: "If you only write code in an air-conditioned room, you will never know what specific problems you will meet on site."
Chuanning's three core fermentation products are erythromycin thiocyanate, cephalosporin and penicillin intermediates. Two years earlier the company told the Shanghai Jiao Tong University laboratory that fermentation metabolism is extremely complex: one production run takes one to four weeks, the final output is under 1% of the feedstock, the tank is a strongly acidic environment with safety risks, and staff had to work 24-hour shifts. Before the AI took over, one 500-tonne tank produced less than 5 tonnes of final product per fermentation.
Li's team moved into the 40-degree workshop in Xinjiang and set up eight desks beside the four- to five-storey tanks as a temporary office. They stayed for 9 months. By generating the fermentation parameter curves for the next 180 hours in advance, with 99.9% accuracy, the model took over the giant tank. Output rose by 3 to 5 percentage points in the month after takeover, according to estimates.
ManuDrive uses a predict, optimise, execute and feedback closed loop on classical control systems. It needs only 5% of the data of a traditional AI model and a dozen ordinary graphics cards, so it can be deployed in a lightweight way. It is delivered through a lifelong service model pairing Frontier Deployment Researchers (FDR) with Frontier Deployment Engineers (FDE), and through a framework of 127 composable modules covering perception, planning, memory and tool calling. A strain metabolism adaptation algorithm built by FDEs for the fermentation project was reused in industrial wastewater treatment by adjusting only 3 parameters, raising cold-start efficiency by 90% and cutting treatment cost per tonne of water by a further 10%.
In equipment manufacturing, an AI process-drawing generator cut single-drawing generation time to under 3 seconds and net profit rose by more than 17%. With Shuangliang Group, design verification per scheme fell to under 10 minutes and net profit rose by more than 15%. If the system covered 30% of China's core fermentation capacity, it would add more than 15 billion yuan a year in industry-wide benefits.
Details