A team of top students from Tsinghua University has created an "AI engineer" to help 20 million engineers build AI assistants | Underwater project

In 2026, the global AI industry reached a critical turning point. The frenzy for large models gradually subsided, and the core narrative shifted from a technological arms race to commercial realization.

In such a cycle, Pinlan Technology, an AI+CAD platform company, chose a relatively alternative development path: it didn't chase the large model trend, didn't create a hit product for the consumer market, and didn't want to tell the story of AI replacing human labor. Instead, it delved into the pile of blueprints and entered the less "sexy" niche of industrial design.

Founded in 2018, Pinlan Technology focuses on the R & D and industrial application of the AI+CAD core platform, providing AI design overall solutions for the AEC engineering design and MFG industrial design industries. The founder, Li Yifan, graduated from the Department of Electronic Engineering at Tsinghua University and later pursued a master's degree at Carnegie Mellon University. During his entrepreneurial period in Silicon Valley, he participated in multiple startup projects, served as a member of the founding team of MailTime and the person - in - charge of industry AI at Caicloud, leading the development of multiple AI - related products and solutions.

In recent years, leading CAD manufacturers such as AutoDesk have also started to deploy AI capabilities. However, Li Yifan doesn't want to be the "Chinese version of AutoDesk." Instead, he aims to build an "AI engineer" that can accept natural - language tasks, understand design intentions and physical laws, and is capable of autonomous planning and continuous self - evolution.

1. Product evolution from blueprint recognition to the "AI engineer"

After Li Yifan returned to China from Silicon Valley in 2018 to found Pinlan Technology, he spent a year looking for application scenarios suitable for AI technology implementation in different industries. It wasn't until 2019 that a speech by Pinlan Technology at a Microsoft forum caught the attention of the Vanke team.

After several months of communication and discussion, Li Yifan learned that architectural design requires a large amount of drawing work. On average, each blueprint needs to be modified 10 times, and repetitive labor accounts for over 80%. Just checking the construction drawings for pipeline layout takes half a month of repeated verification. Meanwhile, a large number of digital blueprints, specifications, and experience precisely provide the core data foundation for AI technology implementation. This made Li Yifan see the possibility of technology implementation, and Pinlan Technology reached a cooperation with Vanke to develop an AI blueprint review system.

Different from simple image recognition, blueprints are far more complex than natural images. They contain highly structured professional expressions such as layer information, professional symbols, dimension markings, and multi - page associations, and there are complex logical relationships among hundreds of blueprints. "Although the so - called 'Four Little Dragons of AI' at that time could do image recognition, as of today, apart from us, no other enterprise can recognize design blueprints," Li Yifan said.

Pinlan Technology's algorithm team started from basic knowledge such as geometric principles and structural mechanics to disassemble the drawing logic, built the underlying model algorithm, and the design department provided industry know - how. The two teams cooperated and complemented each other. From the smallest module of a lightning rod to the electrical system of an entire residential building, and then to the design needs of various non - architectural industries, they gradually built a design model suitable for engineering scenarios that designers can use, thus continuously refining and optimizing the capabilities of Zhuhuitong, the company's core AI CAD product.

The core logic of this technical architecture is to use a general underlying algorithm as the "internal strength" to handle the "moves" required by different industry scenarios.

"Customers don't need to reinvent the wheel. They just need to train their own exclusive models on the platform," Li Yifan explained. As a result, the drawing efficiency has been increased by 10 times, and the rework rate has been reduced by 90%.

Currently, Zhuhuitong has undergone seven key iterations. Based on the technical foundation of large language models, deep - learning image recognition engines, generative design engines, and cloud - based modeling engines, it has the capabilities of cloud CAD, AI image recognition, design large models, and AI design agents. It is now applied to the layout design and blueprint drawing of engineering projects in construction, power engineering, ships, etc.

Pinlan Technology product architecture

Zhuhuitong working interface

"Foreign traditional manufacturers have noticed our products, but the ones more likely to pose a threat to us are actually the large - model giants. However, there are no such signs yet, as large models are still unable to handle design blueprints in this form," Li Yifan said.

2. Shifting from vertical application to a platform - based strategy and going global to Japan in the "ODM" mode

In 2023, the Chinese real - estate industry experienced fluctuations, and the annual development volume of commercial housing dropped sharply, which directly affected Pinlan Technology's original strategy focused on residential AI design.

"The market environment has made us realize a fact. Our advantage lies not in being industry experts but in being AI technology experts. So we should focus on building a good technology platform and leave the industry - related matters to our partners and customers, as they understand better than us," Li Yifan said. "Just like Huawei builds base stations but doesn't sell SIM cards."

Currently, Zhuhuitong is positioned as an "AI+CAD core platform provider," adopting a service model of "software license + implementation service" (License + Service), and it can also co - create and develop customized versions with customers. As of March 2026, the number of Zhuhuitong's B - end customers has exceeded 300, and it has achieved annual operating profitability.
 

Commercial layout

For example, Haier's air - conditioner design application "Haizhuhui" developed based on Zhuhuitong has actually saved 70% of the design time compared to the past. A chain hotel used Pinlan Technology's AI platform to arrange hotel rooms, quickly generating building floor plans and engineering quantity lists that meet brand standards, reducing the original 4 - hour design time to 1 hour, an efficiency improvement of about 75%.

Haizhuhui working interface

Originally, Pinlan Technology didn't have a clear plan for going global. A Japanese customer extended an olive branch of cooperation after seeing 36Kr's report on Pinlan Technology during the process of looking for a supplier.

From receiving the inquiry to officially starting the cooperation, Pinlan Technology spent half a year conducting research and finally chose a co - creation cooperation model: Pinlan Technology provides the technology, and local partners provide local industry experts and sales channels to jointly develop an industry - specific AI design tool for the Japanese market.

This ODM - like path to going global is different from the popular SaaS global expansion in recent years. Instead of directly selling the product after language translation, it requires "redoing everything" from technical adaptation to product definition.

Li Yifan explained, "Our past experience is not to rush to expand on a large scale and have 'awkward conversations' with consumer - end users. Instead, we should first choose a partner with industry know - how for in - depth co - creation and penetrate deeply from a professional direction. Once the pilot is successful and we prove our capabilities to customers, it's just a matter of time before we expand further."

"In highly developed local markets like Japan and South Korea, it must be the locals who do local things." After the cooperation started, Li Yifan quickly realized the "difference" in the Japanese market. Japan's unique architectural form is mainly single - family "detached houses," with fewer centralized apartments, and its design specifications and standard systems are also completely different. Therefore, the model algorithm needs to be comprehensively modified and adapted.

Secondly, the Japanese market has high requirements for product reliability and is more conservative about trial - and - error. It hopes that the launched AI products must be mature enough. "So we have to take our time to refine our products," Li Yifan said. This "refinement" is reflected in all aspects, such as higher seismic design requirements, the model being able to interface with locally commonly used design software, and conforming to the work processes and usage habits of Japanese engineering designers.

To solve these problems, Pinlan Technology adopts a two - location collaborative model: part of the team is stationed in Japan, and part remains in China, and the development is promoted through online collaboration. "Going global can't be rushed. After all, we're not a giant enterprise. First, we need to take a good first step in Japan. In the future, whether we go to South Korea or Europe and the United States, we'll gradually expand from point to area," Li Yifan said.

This "slowness" may be the essence of the engineering technology application track, which requires time to accumulate industry understanding and patience to build trust relationships. When asked about the ultimate vision, Li Yifan said, "Let AI be able to do engineering design and create a better world." This vision is grand, but the implementation path is simple. Starting from the AI design of a lightning rod to the cable system of a nuclear power plant, while others are doing "sexy" things, Pinlan Technology chooses to slowly accumulate internal strength and then wait for time to give an answer.