Hai Zhi Technology: Capital Lessons on Putting "Reins" on Large Models

Over the past three years, AI has been the most closely watched track for global capital. Entering 2026, the capital landscape has undergone a notable shift: investors no longer simply pay for "how large the parameters or the model is", but instead focus their spending on "what the model can actually accomplish".

Recently, Yang Zaifei, CEO of Haizhi Technology, had an in-depth conversation with *Business School* magazine, centered on Haizhi's real-world implementation practices in the AI industry. Facing the persistent "hallucination" challenge of large language models, how has Haizhi built a reliable and controllable solution? What practical experience has it accumulated in the fields of ontology and Harness engineering?

It is hoped that this case analysis will provide a frontline practical reference for understanding the real challenges AI faces as it transitions from technology to industrial application.

01 Putting a "Rein" on Large Language Models

The more intelligent a model becomes, the more it needs to be steered, because the deeper it penetrates into critical business operations, the higher the cost of a single probabilistic error.

Hallucinations in large language models are a widely recognized issue across the industry. In scenarios such as the maintenance of large power grids, the discrete manufacturing of aircraft composed of millions of global components, and regulatory enforcement in the financial sector, even a one-in-ten-thousand risk of a probabilistic, unexplainable outcome can lead to unacceptable consequences. Therefore, the deep integration of large language models into industrial applications requires "de-hallucination". 

This is precisely the origin of the Harness concept that has exploded in the AI field today. Yang Zaifei divides the application of large language models into three progressive stages: The first stage is the Prompt stage, where high-quality queries help the model act as a competent advisor; the second stage is the Context stage, where increasingly extensive context enables the model to comprehend enterprises' more complex knowledge systems; the third stage is the Harness stage, which puts a "rein" on the ever-smarter large language models. 

In essence, a large language model is fundamentally an "inductive model". What it produces is ultimately the "most probable" result, not a "definite" one. Hallucinations are the product of this probabilistic deviation. 

When large language models enter real industrial frontlines, enterprises need "agents" that can truly embed into workflows and independently solve complex problems. "The more intelligent a model is, the more it needs to be steered, because the deeper it goes into critical business, the greater the cost of a single probabilistic error. That's why people propose that the future Agent equals a large language model plus Harness (rein)," Yang Zaifei says. 

Haizhi Technology's method of "de-hallucinating" large language models is to reintroduce deduction in addition to induction — a set of logically traceable and auditable engineering processes. The "graph ontology" is the carrier of this system: it systematically organizes the facts, rules, and interrelationships in the industry into a connected network, and binds the large language model to this network, so that the model checks against the knowledge graph before generating outputs, achieving "graph-model integration". 

Haizhi Technology has accumulated over 10 years of experience in the knowledge graphs used in "graph-model integration", having started its layout back in the era before large language models existed. 

As the technology enterprise in China that has long focused on knowledge graphs, Haizhi Technology has Zheng Weimin, an academician of the Chinese Academy of Engineering, serving as its chief scientist, and has co-established the nation's first high-performance graph computing academician expert workstation. According to Haizhi Technology's prospectus, in the complex logical reasoning and anti-hallucination MuSiQue benchmark test, traditional large language models score below 10%, ordinary vector RAG scores below 40%, while Haizhi Technology's graph-model integration technology has a score exceeding 50%. 

In 2026, two national standards, *Intelligent Computing — Graph Computing Performance Test Methods* and *Information Technology — Neural Network Representation and Model Compression — Part 3: Graph Neural Network*, were successively implemented, with Haizhi Technology being one of the core drafting units for both. The implementation of these two national standards has shifted the competition in the graph intelligence industry from "competing on model parameter size" to "competing on system controllability, governability, and deployability". 

Specifically for Haizhi Technology's Atlas agent, the knowledge graph is embedded into the entire workflow of the large language model: In the pre-training stage, the model first learns the structured logic of the knowledge graph to establish factual cognition in advance; in the reasoning stage, the large language model is responsible for divergent association, while the knowledge graph synchronously conducts logical verification to ensure every conclusion is supported by "entity-relationship" facts; in the output stage, the content is further verified and corrected. Induction ensures speed, breadth, and the ability to draw inferences from one instance, while deduction ensures rigor, controllability, and traceability. The two are deeply integrated, not simply superimposed. 

Currently, Haizhi Technology's platform has integrated over 100 large language models to realize "de-hallucination" for their industrial applications. This is precisely what the capital market values most about Haizhi Technology: as large language models become increasingly intelligent and affordable, capital will reward enterprises that can deploy these models in industries where errors are unaffordable. 

Haizhi Technology's investors include well-known institutions such as BAI Capital, Legend Capital, Hundsun Technologies, GL Ventures, and the Shanghai Artificial Intelligence Fund. Its 2025 annual report shows that the company's total revenue reached 621 million RMB, with Atlas agent solution revenue growing 68.4% year-on-year; net profit increased by 42.6% year-on-year. 

As the lead investor of Haizhi Technology's Series A+ round, Long Yu, Founding and Managing Partner of BAI Capital, stated: "We have witnessed the team's key leap from data intelligence to graph-model integration. As large language models move out of laboratories and into industrial scenarios with extremely low error tolerance, AI hallucinations have become a gap between 'usable' and 'trustworthy'. Haizhi Technology endows AI with the precise reasoning ability of a 'left-brain-like' system, complementing the creative 'right-brain-like' strengths of large language models. It has taken the lead in significantly reducing large language model hallucinations through knowledge graphs, and built technical barriers in three dimensions: factual accuracy, hallucination rate, and complex logical reasoning. Going public is a new starting point for continuous evolution under more transparent and long-term market scrutiny. We look forward to Haizhi Technology turning the industry's consensus on trustworthy AI into China's long-term global competitiveness." 

It is worth noting that since Haizhi Technology can "de-hallucinate" large language models, could its outputs still produce "secondary hallucinations"? Worse, could these "secondary hallucinations" — results that have undergone ontology verification — be disguised under the label of "verified", making errors harder to detect? 

Yang Zaifei says that "de-hallucination" does not mean eliminating hallucinations entirely. "Instead, it keeps the decision-making process within a logically deducible scope that you can track. When the process is traceable, error detection and problem tracing have a basis, and progress becomes measurable." In his view, using auditable methods to address errors is more pragmatic than expecting the system to "never make a mistake". 

02 Distilling Tacit Knowledge into Assets

True enterprise knowledge is often hidden in unstructured data not included in traditional business systems (such as SAP).

In Haizhi Technology's "graph ontology", the knowledge graph connects industry facts into a network, while the ontology is the rule for weaving this network: it defines what types of things exist in an industry, what attributes each type has, and what relationships can be formed between types. The ontology is the "skeleton" and "rulebook" of industry knowledge, and the knowledge graph is the "flesh and blood" that grows from this skeleton. With the ontology, the large language model can understand what is what in an industry and what can be connected to what, ensuring every reasoning step is traceable — this is the prerequisite for "de-hallucination". 

Building an ontology relies on painstaking on-site work, but it is by no means a manpower-intensive tactic. A large amount of industry experience is hidden, and Yang Zaifei divides it into two categories: The first category consists of unstructured operation manuals, on-site notes, and work summaries — content that exists in text but has not been organized into rules. The second category has no text at all, with experience embedded in experts' muscle memory and intuition, which even they cannot articulate clearly. To extract these two types of vague knowledge bit by bit and organize them into an ontology, there are no shortcuts, and one must immerse themselves in the on-site environment. 

However, immersing in the on-site environment does not mean relying on manual extraction bit by bit. Yang Zaifei says that experts' "muscle memory" on-site has long been deposited in the enterprise's digital systems through repeated business operations. True enterprise knowledge is often hidden in unstructured data that is not captured by traditional business systems (such as SAP). What Haizhi Technology aims to do is use AI automation to penetrate the enterprise's digital infrastructure, obtain these know-how and data that cannot be easily captured by large language model vendors, and "distill" them into structured assets reusable by large language models. 

This idea of converting tacit knowledge into structured assets aligns with the world's most cutting-edge AI business theories. 

On June 14, Microsoft CEO Satya Nadella proposed the concept of "Token Capital". He pointedly noted that the purpose of enterprises using AI (consuming Tokens) should not be merely completing tasks, but transforming it into "AI capability assets owned by the enterprise itself". In Nadella's view, precipitatable and reusable enterprise-level AI Harness and knowledge networks accessible to AI represent the true efficiency of the AI era; the key lies in the systematic precipitation of human judgment, experience, relationships, and creativity by AI. 

Yang Zaifei has expressed a similar view: "As large language models move into the deep waters of practical application, the true value of an enterprise does not lie in chasing flashy buzzwords, but in having the ability to dive into the troublesome problem sites and resolve those tedious, labor-intensive tasks." In the capital market and industry, many people compare Haizhi Technology's business model to Palantir, even calling Haizhi Technology "China's Palantir". "Indeed, initially both sides dispatched the smartest engineers to clients' tedious on-site environments. Some people once thought this approach was 'not glamorous', and the ideal state of the software business was to deliver a disc to the client and solve problems from the office," Yang Zaifei says. He believes the path Haizhi Technology has chosen is correct: placing top talent in truly tedious on-site scenarios to derive real conclusions is the most important contribution to society. 

In Yang Zaifei's view, ontology is never a slogan that can be easily shouted, but a painstaking task that heavily relies on industry accumulation. "Many companies started working on ontology retroactively after large language models became popular; while Haizhi Technology has spent over a decade deploying its top engineers to clients' most tedious on-site environments, thoroughly working through those 'slow and meticulous tasks', before naturally distilling this knowledge system," Yang Zaifei says. The real competitive edge does not lie in who first coined the term "ontology", but in who possesses the engineering capability to translate complex industry logic into machine-understandable, executable rules. 

In fact, placing engineers on the frontlines, abstracting on-site experience from different enterprises into knowledge systems, and "distilling" the tacit knowledge that can only be accumulated through deep industry immersion into knowledge that can be reused and generalized by large language models in the future, tests both the organization's deep understanding of industry logic and the team's resolve and patience in AI engineering. 

Yang Zaifei says the ultimate goal of this path is to teach large language models to build ontologies independently. "We hope that when a large language model enters an unfamiliar, in-depth industry, it can identify ontologies, construct ontologies, and discover relationships between ontologies on its own. Through continuous input and interaction, the large language model will reach the fuzzy boundaries of knowledge faster than humans, and the real challenge lies in how humans interact with it to enable it to generate good ontologies in those ambiguous moments." 

Haizhi Technology is working hard on this. For example, in the financial sector, Haizhi Technology is collaborating with several large banks to build "enterprise-wide knowledge engineering", constructing a complete financial ontology architecture to support intelligent auditing, risk control, marketing, anti-money laundering, and anti-fraud. In the energy sector, Haizhi Technology targets challenging tasks such as full lifecycle management of oil wells and rapid fault diagnosis for large power grids. 

Yang Zaifei says many of the directions Haizhi Technology has chosen are in foundational fields. The problems in these fields are largely unsolved, with data that is often highly complex, heterogeneous, mostly unstructured, and not fully grasped by humans or large language models. It is precisely in these fields that greater value can be generated. 

However, as large language models become more capable of self-learning and autonomous decision-making, who will be responsible if they make mistakes in these critical scenarios? Yang Zaifei says that when the system is effectively constrained, auditable, and traceable, the probability of errors decreases, and there is a clear path for tracing after errors occur. "Addressing large language models with proper methodologies is more pragmatic than expecting them to never make a mistake." 

In fact, in fields such as energy and security, humans are not error-free either. But with the ontology as a "rulebook", when a large language model makes a mistake, the system has a path to "trace and learn from errors". 

Regarding the accountability for large language model errors, Yang Zaifei elevates it to the ultimate question of how carbon-based civilization and silicon-based civilization can coexist. "This will eventually happen, testing the very form of our civilization. Some scientists have signed petitions to halt large language model research, but I lean toward pragmatism — development cannot be stopped, and problems can only be solved in the process of development." As for the expectation that a tool should never make a mistake, Yang Zaifei believes it is a delusion that humans can achieve omniscience and omnipotence.

The reporter notes that these foundational fields Yang Zaifei mentioned are also targeted by many ToB AI enterprises, leading to fierce competition. But Yang Zaifei remains very calm: "The fact that many companies are working in this space should be interpreted as good news. If few people take this path, it is often a dead end; many people participating indicates strong demand. As for Haizhi Technology's advantages, this question can only be answered through competition, not by the CEO making pledges. As long as the market is large enough and innovation is strong enough, we will naturally find our own space. Competition energizes everyone, as it is not a zero-sum game, but a win-win scenario that expands the market through innovation." 

*This article is reprinted from the July issue of *Business School* magazine, edited and published with authorization.

This article is from the WeChat official account "Haizhi Technology", written by Zhu Yun, and published with authorization by 36Kr.