DeckSpeed’s Milton Yan: Entrepreneurship Is About Demystifying the World | Gen Z Founders

By Zhu Kexuan

Edited by Chen Caixian

Influenced by his grandfather, Milton Yan immediately thinks of two words when asked to describe his life: engineering and business. He never imagined growing up to work on a factory floor. The seeds of entrepreneurship—and his fascination with products—were already taking root in middle school.

Whenever he had free time, he would follow experienced makers on Baidu Tieba and recreate projects such as electromagnetic launchers, ZVS circuits, and Tesla coils. After repeatedly using templates, modules, and kits bought online to make new things, he realized how much he loved building products.

He enjoyed neatly soldering circuit boards or producing beautiful arcs with a Tesla coil, then photographing the results and posting them online.

Laughing as he recalled his second year of middle school, Yan described combining chemical reagents, connecting an igniter and a remote-control module, and setting off a homemade explosive on the hill behind his school. The resulting mushroom cloud made people on both of the school’s soccer fields stop and watch. His classmates still bring up the incident when introducing him.

That enthusiasm for products continued into college. In his sophomore year, Yan started entering maker competitions and hackathons, earning his first substantial prize: RMB 100,000 for a growing box he had built. Prizes from subsequent competitions added up to several hundred thousand yuan and became his early startup capital.

He attended Babson College, known as America’s leading college for entrepreneurship, expecting to find like-minded people who wanted to start companies. Once there, however, he became disillusioned. “No one in the world can teach you how to be an entrepreneur,” he said.

Later, he would wait outside Olin College of Engineering, pitching his ideas and showing his products to anyone he met in hopes of finding collaborators. He did spend some time building a startup with people from Olin and Harvard. But he eventually lost those illusions too and returned to working alongside teammates he had known for years.

“I think the uncertain, difficult periods we went through helped people who lacked resolve or didn’t truly believe in themselves choose to leave. After raising money, you become more reluctant to give away equity. But when someone committed fully before the funding and kept contributing, you feel they’ve earned it,” Yan said.

“Entrepreneurship is fundamentally a process of demystifying the world.” It was the third time he had used the word “demystification.” Yan eventually left academia and committed fully to entrepreneurship for two reasons.

First, he had secured funding. Second, he felt a sense of urgency: the AI revolution was constantly changing people’s understanding of what was possible, and he did not want to miss the opportunity. He felt he could always return to earn another degree later.

In 2023, Yan joined the ranks of Silicon Valley entrepreneurs. He first built MEEP, a social app sold to clubs. This year, he moved into AI agent infrastructure and launched DeckSpeed, CoreSpeed, and Zypher Agent. Several Silicon Valley startup customers described DeckSpeed as “Cursor for Slide.”

He wants to redefine presentations in this wave of AI agents.

The following is AI Tech Review’s interview with Yan, edited by the original author without changing its meaning.

Entrepreneurship as a process of demystification

AI Tech Review: Why did you want to become an entrepreneur, and when did you decide to go all in? Could you walk us through that journey?

Milton: I never grew up thinking I would look for a job. I wanted to do things that were connected to me personally. Being one of a product’s creators and owners matters a lot to me. Working at a big company would be painful for me because whether the product succeeds or fails would not feel like my own business.

Many people say funding is the signal to start a company. I really began trying entrepreneurship two years ago. Before that, I hadn’t fully committed because I wasn’t confident enough. I only knew that I felt the most passionate when building my own products. What finally pushed me to go all in was that my cofounders and I began to demystify the world. Funding came second.

I think entrepreneurship is fundamentally about demystifying the world. I could never know when I would be ready. We kept discussing how to handle competition and how deep or strong our product’s moat really was. Two years ago, I spoke with some of the world’s leading entrepreneurs and professors and realized that the world is far more improvised than it appears. Nothing is beyond someone’s ability simply because of who they are. From then on, I felt I should do things faster and better, and I committed fully.

AI Tech Review: How many key people are on your founding team? What do your backgrounds have in common, and how do they differ?

Milton: We have five core founding members. The developers are all full time; the others work part time. Last year there were more than ten of us. With the tools available this year, everyone can open two or three Cursor windows and develop in parallel, so we kept only the strongest people. Nearly all of us went to the same undergraduate university. Hao Su, Zi’ang Wu, Yunpeng Zhao, Max, and I have already built many products together.

AI Tech Review: What are their individual strengths and stories?

Milton: Our team won an NVIDIA global hackathon championship. Almost everyone has accumulated six-figure prize money, and we’ve all published papers, though none of us stayed in academia. Hao Su is our CTO and an undergraduate classmate. He started making products in middle school, as did Zi’ang Wu and Yunpeng Zhao. We have a lot in common. When we started our previous company, we pooled RMB 800,000 in prize money. At this stage, everyone is a developer, and I oversee the product.

AI Tech Review: Many Silicon Valley founders set up hacker houses where they live and work together. Do you have one? How does the team collaborate?

Milton: We rented a hacker house in LA and all lived there. We also use Discord to create a virtual office. It can be even more efficient than working together in person. Being physically together can feel too close, while Discord preserves some privacy. We can share multiple screens and work while showing what we’re doing. Each person puts in a full fourteen hours of work a day, excluding chatting, meals, sleep, and commuting. It’s a very efficient way to work online.

AI Tech Review: When exactly did you start building these products?

Milton: We started Zypher Agent this year. DeckSpeed is an application built on Zypher Agent, so the initial development was fast. But we spent another month or two refining it, so the process was by no means easy.

AI Tech Review: Who has influenced and inspired you most as an entrepreneur?

Milton: I have to mention my grandfather. He was born in 1951 and built a radio while still in elementary school. Family circumstances prevented him from taking the university entrance exam, and he worked as a cement worker and an electrician. He finally had the opportunity to take the exam at thirty, then went to university in Zhejiang to study industrial and civil construction. After returning home, he earned money drawing plans and later partnered with local businessmen in real estate. As a child, I saw engineering drawings all over the house and often watched the adults discuss engineering and business. I thought it was very cool. If you ask what kind of man I want to become, those are the two key words: engineering and business.

In high school, I read the Chinese edition of Elon Musk’s biography, titled Silicon Valley’s Iron Man, two or three times. I chose IoT engineering as my university major, then switched to computer science and finance a year later. I built many products in college, both hardware and software.

AI Tech Review: How would you describe your product taste?

Milton: I like simplicity. Our website is black and white because we make a design tool; other colors could interfere with users’ aesthetic judgment. I like the art of negative space and particularly dislike visually cluttered, all-in-one interfaces. Although I’m used to IDEs such as Visual Studio, I still don’t find those interfaces beautiful.

AI Tech Review: Why did you choose office productivity as your use case?

Milton: I wasn’t choosing a sector so much as responding to an obvious market need. Documents and office work are essential, and the market is large. But when making presentations myself, I couldn’t find an AI presentation tool that I could comfortably trust to produce something ready to deliver. Nor could I find one that let me express everything in natural language through an interaction model like Cursor’s. That’s why I decided to build this.

AI Tech Review: Could you introduce DeckSpeed from a founder’s perspective? What are its distinctive features and innovations compared with similar products?

Milton: Some products claim to generate presentations, but the output looks obviously AI-generated and isn’t ready to deliver. You can tell at a glance which product made it. Our principle with DeckSpeed and CoreSpeed is to make something people genuinely want to use for a finished deliverable.

We’re operating on a different level from some AI presentation tools. First, our implementation is different. Second, they produce conventional deliverables, while we avoid templates and explore a presentation format with possibilities from zero to one and beyond. Templates accommodate existing workflows; they don’t define a new way of interacting. I don’t really want to compare ourselves with presentation tools. I’m trying to redefine presentations. Underneath is a coding agent, so people more often compare us with Cursor and call us “Cursor for Slide.”

AI Tech Review: Many products really do put the template before the thinking.

Milton: Exactly. Accommodating a template often makes people forget what they originally wanted to express. I really dislike being constrained. Expression should come first, instead of compromising until the final result diverges from the original idea. Our website has a line about making presentations like a boss. I generally don’t want to waste time making slides, but some things I have to do myself because a designer may miss the main point and emphasize something unimportant for visual appeal. That doesn’t work. I want an agent to let me focus on expression.

Building Cursor for presentations

AI Tech Review: In the two years since ChatGPT became popular, which AI-native products have you admired most?

Milton: Cursor is the first one that comes to mind. We’re all heavy users, and it gave us some product inspiration. CoreSpeed solves backend integration problems for Cursor users. DeckSpeed’s interaction model is just like Cursor’s. At its core, DeckSpeed is Zypher, a coding agent, but its output is a presentation.

One day, I used Cursor to build a webpage, converted it to a PDF, and used that as our business plan. But Cursor looks like an IDE, much like Visual Studio, and is designed for programmers. You can’t expect everyone to use Cursor to make slides. So we built something that covers the code with a cloth and draws a presentation on top.

Then there are mobile apps like Cal.ai, which wrap models and do two things well: lower the barrier to using them, and manage the data created afterward. That is distinctive value, and these products have demonstrated strong revenue.

AI Tech Review: There still aren’t that many truly good products. For instance, there isn’t even a one-click tool to convert a Chinese résumé into English across formats and languages. What do you think?

Milton: Some developers in the Chinese-speaking community on X have insight and efficiency far beyond many American teams. They may live in China but publish their products and opinions only on Twitter. From what I see, they barely sleep. Their work is both cutting-edge and substantial.

A language model alone can’t translate a document from Chinese into English while keeping its formatting unchanged. That needs an agent product combining purpose-built tools. Language models expand the boundaries of capability, but the work needs to be divided among specialized tools—for reasoning, coding, translation, and so on—instead of asking one tool to do everything. Hallucinations can also cause unintended changes: you ask for translation, and the formatting gets changed too. That’s not acceptable.

Before deciding what to do, an agent should determine whether the task concerns formatting or content. If it’s formatting, there’s no reason to invoke a content-editing tool. That is how you get consistent output. DeckSpeed can do this. If a small ball in an animated chart is blue, one prompt can turn it pink without changing its shape. Or if I want a line drawing on slide six, it generates one and places it appropriately. One part of the process handles drawing, while another agent checks the structure.

AI Tech Review: Making presentations requires a fair amount of input. Do you currently support voice input?

Milton: Not yet, but that part isn’t difficult; we just need to add text-to-speech. The fine-grained requests that need spoken expression may come later in the presentation-making process. Early on, you can simply provide a website and ask for charts and visualizations.

AI Tech Review: Some argue that you can’t build a good agent without a foundation model and reinforcement learning. Do you agree?

Milton: Technically, having your own foundation model and reinforcement-learning alignment can significantly improve response consistency and stability across multiple turns. That argument holds particularly for products involving complex task decomposition, long-term memory, and strategy selection. Recently, many products have called themselves general-purpose agents, but if a foundation model can already do all of that, you may not need a separate general-purpose agent.

From an engineering perspective, though, an agent is a product of architecture and strategy. DeckSpeed focuses on using tool calls and task scheduling to organize a system with long-term memory and intermediate goals. Berkeley published a paper two weeks ago arguing that language models already have broad, powerful capabilities, but often fail to coordinate well in particular situations. Fundamentally, it’s an organizational problem, much like disorganization among humans. You can’t ask a reasoning agent to handle execution. There is room to improve not only interaction, but also the application layer and agent scheduling, which can determine the ceiling of a product’s capabilities.

When GPT-4o’s image-generation model was announced, I decided we should immediately prepare and integrate the 4o API to use it as our main image generator for presentations. We kept that work off the product’s main branch. Once OpenAI made the image-generation capability available through the API, we could merge the branch and make it available in DeckSpeed the next day. I think that is the right direction. Generating an image separately still forces users to switch tools, copy and paste, and figure out where it belongs on a slide. That is the difficult part.

AI Tech Review: How has MCP affected your agent development? What important changes came with it?

Milton: DeckSpeed was the first slide agent to support MCP. Once it did, a presentation stopped being merely a presentation. Every AI tool on the market claims it can generate slides, but in many cases it is just a workflow: inserting a language model into an existing office process or stitching templates together. Those approaches address isolated needs without really supporting structured creation. DeckSpeed takes on something harder and more valuable. Instead of uploading a document to generate slides, you start with an idea and progressively organize it into a clear structure of expression. The fundamental problem is how to explain a set of ideas clearly: moving from unstructured content to clarity, and from individual writing to collaboration with agents. We’re building an AI-powered system for expression, not just a slide generator.

We have a product called ZhiZhi at zhizhi.corespeed.io. Traditionally, you upload a presentation or PDF, and it turns it into a podcast. With MCP, after making slides through natural-language conversation in DeckSpeed, you can ask it to add podcast narration. DeckSpeed automatically calls the MCP server for ZhiZhi’s podcast-generation service, turning the presentation into a combination of narration and visuals. That redefines what a presentation can be.

Another useful application is that MCP gives you access not only to external tools but also to data you couldn’t otherwise reach. Take Linear, a professional engineering and project-management product. You can publish individual tasks and have employees or developers claim them. At the end of the week, you can see who completed which tasks.

For a weekly report, for example, Linear has an official MCP server. Once users connect it to DeckSpeed, they can retrieve Linear data with a sentence and generate a report. The report can include charts, such as a pie chart showing the breakdown of work or a line chart showing working-time trends. MCP opens up enormous possibilities for weekly reporting and presentations. The output is no longer just a conventional arrangement of text and images.

AI Tech Review: A recent post raised concerns about agent safety and stability. A developer said an agent operating their computer deleted important information. How do you think about protecting critical files in office workflows?

Milton: I don’t think we can blame the agent itself for safety and reliability problems. The question is whether its designers provided enough safeguards. We added an Error Detector. One rule is that it must not touch things it shouldn’t. High-risk operations such as deleting data prompt the user to confirm again.

DeckSpeed also has version-control interactions called Accept and Reject. After giving an instruction, making a change, and seeing the result, the user can accept it or roll back to the previous version and rewrite the prompt. These safety issues can be avoided.

Gen Z is setting the direction

AI Tech Review: What is your core moat?

Milton: We’re building a moat through accumulated data and users. In choosing a product direction, we look for compounding value from tools. A single agent may not be remarkable, but combining two or more tools to produce an impressive deliverable could be the right direction. Most importantly, there may be a hundred combinations for accomplishing the same task. Insight and rapid iteration let you establish a position quickly. For user growth, we can learn from Cursor. Its popularity came partly from people showing what they had made with it. We want to take the same route and let people see how good the output can be.

AI Tech Review: What are your next plans for DeckSpeed?

Milton: I want people to explore more use cases. A well-known consulting firm has already adopted the product. Many employees in the US spend every day making presentations, and our tool can produce genuinely deliverable work at the level of research reports. They use it to reduce employee hours. Even if we charge large enterprises later, we’ll start with small businesses, then consumers. I plan to keep exploring for a while.

AI Tech Review: How much funding have you raised? Any interesting stories from meeting investors?

Milton: We’ve raised US$1.2 million so far. CoreSpeed Inc. started at the beginning of this year, with funding from the US and Hong Kong.

I prefer investors who understand our products. I don’t especially want them to focus on anything beyond the product and the project. People drawn by prestige will leave disappointed because we’re not a team with particularly glamorous credentials. Ideally, they discover the product first and the team behind it afterward.

Our Hong Kong investor approached us. He was already a partner, saw CoreSpeed online, and invested after a conversation. Another investor at a US-dollar fund was a friend of my school president; they’d played golf together for fifteen years. We spoke on the phone for forty minutes, and he decided to invest without meeting me in person.

I think the way you’re introduced matters. If you approach someone directly, they may start out guarded. If a mutual acquaintance knows you well, you’re a second-degree connection, and you can meet with more openness and trust. When they understand and believe in what you’re doing, you have more agency in the relationship.

ZhenFund previously announced a RMB 100 million initiative for founders born after 2000. Seventeen founders were selected, and I was one of them. While they were still discussing internally how to distribute the money, all seventeen of us were gradually raising funding elsewhere.

AI Tech Review: As a founder born after 2000, what new changes do you think you can bring to technology entrepreneurship, investors, and users?

Milton: It’s something fundamental. Everyone ages. As physical energy declines, people may become less receptive to new things. Young people’s lifestyles and openness lead them to advocate different things from older generations, which changes their approaches and ideas. Experienced people know how to read the signs, but by the time they see those signs, someone young has already created them. That’s a huge advantage.

AI Tech Review: What has been your strongest feeling over these two years of entrepreneurship?

Milton: It’s exhilarating. There is occasional pressure, but I’m happy every day. I don’t have a favorite leisure activity because I’m already doing what I love. The best moment is the night before a product launch. We imagine how well people will respond, tell each other it will be a huge hit and beat everything else, and get carried away with excitement. If we can feel that way, it also means we’ve made something really good.

AI Tech Review: How do you maintain your startup’s pace while learning new technologies and keeping your knowledge current?

Milton: We keep working on the newest things. AI presentations may be a familiar topic, but MCP was released only two months ago. As soon as it came out, we built CoreSpeed on it, followed by Zypher Agent and DeckSpeed, then made DeckSpeed the first slide agent to support MCP. We keep working at the frontier and hope that advances in foundation-model capabilities will help us improve our products.

Working on the front line of AI applications naturally keeps our thinking fresh. These developments directly affect us. I’m not forcing myself to study just to avoid falling behind. New capabilities can make my products better, so I have a direct stake in learning them. Leiphone (WeChat public account: 雷峰网).

Original article by Leiphone. Republication without authorization is prohibited. See the republication guidelines.