Most industrial AI projects fail to deliver measurable returns on investment, and Zhi'an Antu aims to build the "acceptance layer".
Over the past two years, capital has flooded into the industrial AI sector at a rapid pace.
According to IDC data, the proportion of Chinese industrial enterprises applying large models and intelligent agents rose to 47.5% in 2025, up from 9.6% in the previous year; according to estimates by 36Kr Research Institute, the implementable market size of industrial AI in China will reach about 116.3 billion yuan in 2026, and will exceed 300 billion yuan by 2030.
Another set of figures paints the opposite picture. The 2026 Industrial AI Readiness Survey shows that 68% of manufacturing enterprises are still stuck in the pilot, proof-of-concept or pure research stage, and only 7% have truly integrated AI into their core processes. A 2025 MIT study is more straightforward: 95% of corporate generative AI projects have failed to generate measurable financial returns.
Money has poured in, but the returns cannot be calculated.
In the *2026 China Industrial AI Industry Development Research Report*, 36Kr Research Institute puts forward a judgment that the key to breaking the deadlock lies in "scenario-based adaptation and closed-loop verification". In other words, the bottleneck is not whether the model is powerful enough, but that no one verifies whether a modification after adjustment is valid or not.
What Hi-Tech AnTu (Beijing) Intelligent Equipment Co., Ltd. aims to do is exactly to solve this problem.
I. Stuck at the last mile: AI enters the workshop but not the process
The demands raised by industrial customers are often not the same as what software companies understand.
The phrase "too many false positives" translates into engineering language as: how many alarms were manually ignored in the past month, what types of alarms were ignored, and whether real incidents were missed after the ignorance. Without answers to these three questions, modifying the model is purely based on feeling. However, on industrial sites, decisions can never be made by feeling — as one interviewed engineer put it, if AI produces instructional errors, it will cause an irreversible systematic risk to the production links.
As a result, a very common scenario emerges: the pilot test is successfully completed and recognized by the boss, but six months later, the system is still only running on that single production line. The industry estimates that the average sunk cost of each failed scaling attempt is 2.3 million US dollars.
Hi-Tech AnTu's judgment is that what is missing here is not algorithms, but three types of infrastructure: the basis for modification, the verification of modification, and the traceability of modification.
Their first landing battlefield is not the production line, but the construction site — work safety is the segment of industrial AI with the most solid budget and the most likely to be driven by policies.
One of the delivered scenarios is a municipal engineering construction unit in Beijing. The cameras on the construction site were originally only used to review videos afterwards for accountability checks; what they aim to do is to make the cameras give early warnings — to check whether workers are wearing safety helmets, reflective vests and safety belts, whether there are protective measures for edge operations, and whether people break into hazardous areas. These tasks sound like recognition problems, but the real difficulty lies in the judgment criteria: what counts as "not wearing correctly", how long of staying counts as "trespassing", and whether two sets of standards should be applied for day shifts and night shifts. They checked 24 rules one by one with the client's person in charge of safety, and finally worked out a set of solutions that the site is willing to use.
The same process was repeated in the catering chain and cold storage supply chain scenarios, with 24 rules and 20 rules respectively. The three industries are vastly different, but they share one common point: "being able to understand the image" is a general capability, and what really determines whether the solution can be put into practical use is the judgment criteria. These criteria cannot be set by engineers, but must be confirmed item by item by the client's experts. This method of "experts setting criteria and the system executing" is a universal solution they have verified across the three industries.
II. MEDC Engine: No improvement of model capabilities, only management of "improvement"
Hi-Tech AnTu's technical base is called MEDC Engine, whose full name is Evolution Control Plane.
The name itself defines its boundary: it is neither an Agent framework nor execution orchestration. How the business runs and how tools are adjusted are the matters of the execution plane; what MEDC manages is another thing — which experience is worth retaining, which modifications can be launched online, and which improvement is truly effective.
It consists of two sets of models running separately. One set works online, responsible for single judgment completed within 0.3 seconds, outputting a calibrated probability instead of a generated text. The benefit of this design is structural: it is architecturally impossible to "fabricate a non-existent detection", because the system does not generate content at all. When the image is unclear, it will abstain natively, preferring to miss reports rather than make false positives.
The other set works offline, extracting judgment criteria from the execution tracks — why this judgment was correct, why that judgment was wrong, and where the boundary lies. It does not participate in real-time judgment, but is only responsible for converting what happened on site into reusable constraints for the next run.
There is a two-way interaction between the two sets of models: experience improves decision-making, and execution evidence flows back. There is an access control mechanism between them with six steps: improvement proposal, isolated evaluation, three inspections of safety, generalization and regression, risk approval, gray release, and real result verification. One principle runs through the whole process: unverified experience and improvement cannot be automatically imported into the production environment.
The effect of this mechanism can be partially seen from their measured data. A single camera generates 86,400 frames of images per day. After classification, 23.9% of the frames are judged as empty frames and discarded directly, 43.9% are blocked because the shot is incomplete or the quality does not meet the standard, 15.2% are directly output as conclusions with high confidence, and only 6.3% need to be judged by the cloud large model. About 87% of the images are processed locally. This set of data comes from a deliberately selected test set of hard cases (nighttime, occlusion, distant small targets), and the empty frame rate of real fixed cameras will be even higher.
This is not just about saving costs. What industrial customers care most about is that data does not leave the factory — in industries such as energy, military industry and semiconductors, images are not allowed to leave the park at all. Therefore, their products are made in three forms: cloud API/MCP, for teams with R&D capabilities to call on a pay-as-you-go basis; privatized SDK, for financial, government, central enterprises and state-owned enterprises to integrate into internal networks; end-side hardware MEDC Box, with 32 channels per unit and available when disconnected from the network.
The three forms share the same core, with the only difference in deployment locations. The prices are also classified by form: API is billed by call volume, SDK is sold on an annual basis, hardware is quoted per unit, with additional implementation and annual subscription services.
The customer profile is therefore quite clear: the first category is large and medium-sized enterprises that have their own AI teams but lack verification and access control capabilities; the second category is institutions restricted by regulations that do not allow data to leave the factory area; the third category is system integrators and equipment manufacturers that want to integrate visual capabilities into their own products.
III. A 9-person company, why dares to build a platform
Frankly speaking, this is not a light model.
Hi-Tech AnTu currently has 9 employees, and plans to expand to 21 people within 18 months, most of whom are engaged in engine R&D and industry delivery.
Founder Li Xin has 8 years of experience in deep learning inference application architecture, led the team to independently develop multiple industrial-grade quality inspection systems, robot intelligent assembly systems, and also developed multi-agent large models. This resume explains their priority: the same team can choose to build a more powerful model, or choose to solve the engineering problems in model landing — inference latency, end-side computing power, false positive rate, and judgment criteria. They chose the latter.
"Safety-related control should always be undertaken by certified DCS, SIS and PLC, and we will not touch that part," Li Xin said, "We only answer another question: what needs human review, what can be launched online, and which improvement is valid."
The company has set a principle for personnel employment: "do not take projects that cannot leave valuable assets". After each order is delivered, three things must be retained: rule base, evaluation set, and failure list. They would rather reject the project if these three things cannot be retained. At present, 2 patents and 6 software copyrights are under application.
This also explains why their revenue structure seems a little "impure". At present, the two application-layer products that actually contribute cash flow are: Antu Zhiwei for physical safety monitoring, and the knowledge compilation agent for enterprise knowledge governance. According to the 2026 forecast, these two segments together with training and operation and maintenance will contribute about 76% of the revenue.
But in their account book, these two are "derivative products", positioned as the first batch of applications of the engine — to verify boundaries, accumulate industrial assets, and support current cash flow. The real core products are the three delivery forms mentioned above. According to the plan, the proportion of core product revenue will rise from 24% in 2026 to 50% in 2028.
There are four things supporting this growth slope, which they believe can only be accumulated over time: executable test cases covering industrial failure modes, failure case library archived by industrial category, on-site tracks, and co-signed customers who are willing to put their names in the contract. The goals are to reach 60 categories, 600 cases, 20,000 tracks and 8 customers respectively within 18 months. The first version of the baseline will be released 90 days after the financing, and subsequent updates will be disclosed quarterly.
Financially, their plan is to achieve a revenue of 10.84 million yuan in 2026, 40.01 million yuan in 2027, 111 million yuan in 2028, and turn operating profit positive in 2028.
Hi-Tech AnTu is launching its Pre-A round of financing, planning to raise 10 million RMB, transfer 10% equity, with a post-money valuation of 100 million yuan. 43% of the funds will be used for team expansion, 22% for evaluation assets and product R&D, 18% for market and agency system, and the rest for computing power reserve and FDE delivery team.
In this track, the question more difficult to answer than "building a more powerful model" is: why customers can trust that your modification this time is correct. This 9-person company is betting on the latter question.
About Hi-Tech AnTu
Hi-Tech AnTu (Beijing) Intelligent Equipment Co., Ltd. was founded in December 2025, located in Zhongguancun (West Beijing) AI Science Park, Beijing. Its core product MEDC Engine (Evolution Control Plane) is delivered in three forms: API/MCP, SDK and end-side hardware; its derivative products are Antu Zhiwei and the knowledge compilation agent, which have been delivered in three industries: construction sites, catering chains and cold storage supply chains. The company currently has 9 employees, with 2 patents and 6 software copyrights under application. In this round, it plans to raise 10 million RMB and transfer 10% of its equity.