Kando AI has closed a seed round financing of tens of millions of yuan, aiming to build "the Cursor in the decision-making field" | Emerging New Project

Article by Deng Yongyi

Edited by Zhang Yuxin

Wu Bingzhe conducts at least one post-analysis review every day.

As a Doctor of Computer Science from Peking University, he is also a high-frequency investor. After market close each day, he reviews his judgments from that day—what contingency plans were executed, which were disrupted by new market information, and which subconscious decisions were later proven correct.

This habit has lasted for many years. However, one problem remains: most of these post-analysis reviews have not been systematically documented.

"Your cognition is an asset, but right now everything is stored only in the human mind," Wu Bingzhe says. He and Mao Shuhan, CEO of Kando AI, are both active secondary market investors. They share the common trait of frequently making decisions in uncertain environments, and they have both realized that the cognitive patterns behind these decisions have never been digitized.

In June 2026, the two officially founded Kando AI, aiming to enable AI to continuously learn from real-world decisions and their outcomes, and build the next-generation self-evolving decision-making system.

According to "Intelligent Emergence", Kando AI has completed a seed financing round of tens of millions of RMB, led by StarLian Capital, with participation from Force United Financial Holdings and well-known industrial investors. The Kando team currently has around 10 members, coming from leading domestic institutions such as Tencent, ByteDance, Meituan, and Huatai Securities. Founder and Chief Scientist Wu Bingzhe graduated from Peking University with a bachelor's degree in Mathematics and a doctorate in Computer Science. He was the first recipient of the Apple PhD Scholarship in mainland China and the former head of Trustworthy AI at Tencent AI Lab. Co-founder and CEO Mao Shuhan is a serial entrepreneur, who earned his bachelor's degree from Tsinghua University and a master's degree in Finance from the University of Hong Kong. He worked at a leading domestic investment bank for many years, co-founded an embodied intelligence company, and was named to the 2025 Forbes 30 Under 30 list.

From Information Digitization to AI-Powered Decision-Making

What Kando AI aims to build is not just a tool, but an environment where cognition can continuously evolve. To sum up what Kando AI is doing in one sentence, CEO Mao Shuhan puts it this way: "In the era of the Internet and mobile Internet, we completed the digitization of information; in the AI era, what we want to promote is the self-evolution of decision-making."

Over the past two decades, whether it's financial platforms like TongHuaShun and Wind, or academic research tools such as paper databases, literature management tools, and knowledge platforms, the problems they essentially solve are all at the "information layer": helping you collect, organize, store, retrieve, and summarize information. However, the organized information, stored somewhere, does not naturally turn into valuable judgments.

Whether in investment, scientific research, or other high-value decision-making scenarios, the truly critical step often takes place in the human mind. Researchers form judgments from massive amounts of materials, secondary market investors develop a "market sense" amid complex market noise, and scientific researchers build their own aesthetic standards, intuition, and trade-off preferences for a certain research direction through long-term accumulation... These things are extremely difficult to express directly, let alone be simply digitized.

Mao Shuhan believes that there has always been a missing bridge in this process.

"Information and actions are not automatically connected. The most critical link in between is decision-making, and human brains, experience, subconsciousness, aesthetic preferences, and judgment habits have never been systematically preserved in the past. For the first time, AI makes it possible to build this bridge."

This is also the biggest difference between Kando AI and most other AI tools: many AI products still focus on "optimizing output quality"—making summaries more complete, searches faster, and writing more fluent; while what Kando AI wants to optimize is something else: the decision adoption rate and post-hoc reliability, that is, whether users will truly adopt the suggestions given by AI, and whether these suggestions are actually integrated into users' research, judgment, and action processes.

Mao Shuhan states that Kando is positioned as the Cursor in the decision-making domain.

When AI coding tools first emerged in 2022, the code adoption rate was almost zero. Today, Cursor's adoption rate has reached 60%-70%.

The reason for choosing finance as one of the validation scenarios is that this field is one of the most structured domains in terms of information dimensions. Mao Shuhan explains that the financial scenario scores highest comprehensively in aspects such as feedback clarity, feedback cycle (at the day/week level), auditability, and decision value, making it one of the best fields to build a self-evolving decision-making engine.

Turning the Product into a Continuous Post-Training Process

Kando AI's product takes the form of a workbench for high-value decision-making scenarios. But from the team's perspective, the workbench is just the product entry point, and what it truly carries is a complete decision-making system that can accumulate experience from real decisions, validate judgments, and continuously evolve.

One core capability of Kando AI is "memory". However, Kando AI's memory is not a knowledge base in the ordinary sense, nor is it the common RAG retrieval.

Traditional RAG is more oriented toward semantic similarity matching, while Kando AI's memory system is more like a recommendation and recall mechanism for cognitive structures—what the system remembers is not just what users have viewed, but also how users understand problems, their preferences, and how they revised their judgments in the past, more like a Harness.

"It will understand you better and better, but more importantly, it will learn from every interaction between you and the real world, acting as your partner."

Taking the scientific research scenario as an example, the product will generate predictions, and users can choose to adopt, modify, reject, or include them in tracking. After the actual outcome occurs, the system will go back to the initial assumptions, analyze which judgments are verified, which variables have changed, and whether the errors come from missing information, reasoning deviations, or execution problems.

These decision trajectories will not stay in chat records. Users' adoptions, modifications, and rejections, the gaps between expectations and actual results, as well as post-analysis reviews and attributions, will be organized into new learning signals, continuously updating the system's long-term memory, task strategies, skill scheduling, and model adaptation layer.

This means that the outcomes generated by the previous round of decisions will change how the system recalls information, organizes reasoning, and generates suggestions in the next round, forming a recursive closed loop of "judgment—action—feedback—update".

From Kando AI's perspective, this represents an important change for AI products: in the past, model capabilities were mainly determined at the moment training was completed, while in the future, the boundaries of intelligence will increasingly be determined by the experience accumulated after deployment.

Wu Bingzhe defines what they are doing as a kind of "generalized post-training".

In the traditional sense, post-training is more about adjusting model parameters; but from Kando AI's perspective, what truly needs continuous optimization goes far beyond parameters themselves, including how memory is stored, how to perform fine-grained recall, how to forget information, how to design product interaction logic, how the context of skills dynamically evolves, and how different users' feedback can be organized into effective learning signals.

"You can regard the process of building the product as a model post-training process. The only difference is that in the past we optimized parameters, but now we optimize context, interactions, and memory," Wu Bingzhe says.

This is also why Kando AI chose to develop its entire infrastructure in-house. Wu Bingzhe explains that Kando AI wants all training pipelines, feedback pipelines, and memory pipelines to be as transparent, monitorable, and traceable as possible. What Kando AI accumulates is not a superficial layer of product features, but a complete data flywheel that continuously grows around the real decision-making process.

In terms of commercialization, Kando AI chooses to start with high-value, high-frequency decision-making users. Its current seed users are mainly professional groups, including investors and scientific researchers.

Kando AI is now in closed beta, providing services around users' continuous usage, continuous memory, and feedback, covering vertical industries such as finance and scientific research. From the team's perspective, in the future, any decision-making field that is high-value, non-standardized, and highly cognition-dependent may become an application target of this system.

"Kando AI is not just building a tool. In those fields where there are no standard answers but are most worth amplifying, we will truly open up the closed loop of 'information→decision→feedback'," Wu Bingzhe says.

This is probably the most accurate positioning for Kando AI: what Kando AI hopes to eventually build is a decision-making intelligent infrastructure that can continuously absorb human experience, understand real-world feedback, and continuously improve its own capabilities through every decision made.

Li Wenjue, Partner at StarLian Capital, states: The next stage of AI applications is not just processing information more efficiently, but entering real workflows, learning how professional users make judgments, take actions, and revise their cognition. Kando AI chooses high-value decision-making scenarios such as finance and scientific research, takes personalized memory and cognitive modeling as its foundation, and converts users' research processes, adoptions and rejections, and post-analysis reviews into learning signals. It breaks through the boundary of one-off answers from traditional AI tools, forming a decision-making system that evolves together with users. We value the team's combined expertise in cutting-edge AI research, model post-training, system engineering, and complex decision-making experience, and we also value its potential to accumulate proprietary workflow data from real usage and drive continuous capability growth. If this closed loop is verified, what Kando builds will not just be an investment research or scientific research tool, but possibly a new-generation cognitive infrastructure that connects information, judgments, and actions, allowing the most valuable judgment capabilities of individuals and organizations to be preserved, calibrated, and reused.

Image Source | AI Generated

Feel free to reach out for discussions

This article is from the WeChat official account "Intelligent Emergence", author: Deng Yongyi, published with authorization from 36Kr.