A temporal-control large model called ManuDrive, developed by Shanghai Jiao Tong University's AIMS Lab, is pushing LLMs beyond text generation into running production lines, Caixin-style financial outlet CLS reported. Dubbed the "AI veteran engineer," it has deployed at dozens of listed companies and generated over 100 million yuan in measured economic gains for clients.
The starting point was biotech fermentation. The team spent nine months on-site at Chuan Ning Biotech in Xinjiang on data, training and deployment; the company's 2025 annual report notes that with identical material inputs, fermenters run by the AI engineer beat control groups by 3-5% in average yield. Founder Li Jinjin says the model forecasts the full 180-hour fermentation curve at the batch's outset with 99.9% accuracy; taking over a 500-cubic-meter fermenter lifted yield about 5% while cutting variability by 50%.
The difference from veteran operators lies in control strategy. When pH drifts low, the traditional fix — cooling or cutting feed — costs yield; the model fine-tunes sugar-feeding parameters instead, stabilizing pH without sacrificing output. Most AI handles static tasks, but production lines evolve continuously and control decisions hinge on historical state and feedback, which is precisely where temporal-control models come in.
The same base has spawned multiple "AI engineers": a CAD drawing generator built with Taiyuan Heavy Industry produces a drawing in under 3 seconds, lifting net profit by more than 17%; a structural design agent with Shuangliang cut steel-tower design-and-validation cycles to under 10 minutes, raising efficiency over 85%. The platform's modular agent framework has 127 composable components — algorithms packaged for fermentation transfer to wastewater treatment by tuning three parameters, with cold-start efficiency up 90%.
Commercially, Jincheng Technology has entered revenue-sharing with clients and expanded through partnerships with local state-owned capital into ten-plus provinces including Shanxi, Inner Mongolia and Shandong, prioritizing heavy-industry scenarios that replace large amounts of human decision-making.
From fermenters to drawing shops, ManuDrive shows LLMs can enter production control itself; the privacy of line data and depth of scenario binding are becoming the new competitive moat.