Flowtica Rian: The old rules of hardware haven't changed; the new opportunity is doing Memory well

The gimbal team DJI wanted to acquire has built four hardware categories over ten years and is now betting on AI Memory.

Author: Wu You Editor: Dong Zibo

After Plaud went viral, a Chinese company called Flowtica also burst into public view in a short time. In the first half of last year, it launched the world's first AI pen, the Flowtica Scribe, on Kickstarter, ultimately raising $330,000. The pen weighs only 34.5 grams, is equipped with a studio-grade MEMS microphone and a 280mAh battery, supports up to about 30 hours of continuous recording, and integrates recording, transcription, and summarization. What is little known is that Flowtica's founding team did not start from zero. Founder Rian previously had more than ten years of hardware entrepreneurship experience, from handheld gimbals to agricultural sorting robots. The engineering foundation this team has accumulated in machine vision and edge AI is now all being staked on the new track of Agent-native hardware, and Flowtica is the answer they have delivered. Slightly different from the "DJI ecosystem" profile favored by VCs last year, Rian did not leave DJI to start a business, but his growth path has deep ties to DJI. Born in 1992, Rian was Professor Li Zexiang's last academic graduate student. Together with several classmates, he represented HKUST and achieved the best result in the history of the Robocon competition, a record that remains unbroken. At 22, Rian started his first business making handheld gimbals, taking the lead in achieving phone control and extreme stabilization, and nearly being acquired by DJI. His second startup, a smart litter box, reached $1.3 million in crowdfunding. Turning again, he made agricultural fruit sorting equipment, which quickly rose to second place in the industry. What today's AI hardware entrepreneurs are experiencing, and the milestone results they have achieved, Rian has already been through once, and these results are even the reason he has repeatedly pivoted: "We predict the endgame of competition in a pure hardware business: the lead period is only six months to a year. Once two giants stand side by side, scarcity disappears. After latecomers catch up in product capability, they only need to offer equivalent quality at 60% to 70% of the price, and a price war becomes inevitable. Rather than falling into commercial competition over operations, we want to create something differentiated and new." To use a currently popular term, Rian before founding Flowtica was more like a Builder. He is good at feeling products from the user's perspective. Once he finds that even the best solution on the market falls short of expectations, the desire to create is triggered, using the most advanced technical capabilities to raise the product user experience by a level—the extreme stabilization and panoramic stitching of handheld gimbals, the front-back tilt structure of the smart litter box, the precise recognition of the fruit sorting robot... Founding Flowtica feels more like a return for Rian. After on-device large models and Agent capabilities matured, the hardware forms he once wanted to build but was constrained by model capabilities can finally be made. After launching the recording pen, Flowtica immediately launched the new-generation Flowtica Verso, a smart hardware device that can magnetically attach to the back of a phone, features an e-ink screen, and serves as a portable physical entry point for an Agent that both records and executes. Before Verso opened overseas pre-orders, Leifeng.com · Jingxi came to the Flowtica office and talked with Rian and his team, reviewing Rian's more than ten years of entrepreneurial experience and what he is currently doing. This article mainly answers three questions: 1. Over the ten years of making hardware across different tracks, what methodologies has Rian developed? 2. In the Agent era, Rian's judgment on the trends of AI hardware; 3. With competition in AI recording and memory hardware so fierce, what answer has Flowtica delivered? 01 Hardware startups that only compete on specs will inevitably end in a price war

Jingxi: Your entrepreneurial experience is very rich. Could you share how you became involved in hardware entrepreneurship? Rian: I studied at HKUST for my undergraduate degree and was relatively fortunate to meet Professor Li Zexiang. I was Professor Li's last academic graduate student. At the time, I also followed him to compete in the RoboCon robot competition. We were the last cohort he led in the competition and achieved the best result in HKUST's history, and also the best result to date. Let me share one thing that left a deep impression on me. When DJI released its first-generation aircraft, the price of the Phantom 1 to Phantom 2 increased, yet sales surged instead. That shocked me greatly at the time—a Chinese hardware brand raised its price through product strength, and users bought even more. Later, DJI drove 3D Robotics out of the ToC market, which shocked us even more. We truly felt the strength of Chinese intelligent manufacturing and hard-tech innovation, and it also built some confidence. It was also in those years that Professor Li gradually felt that his best students should no longer be sent to Silicon Valley, and began keeping students to start businesses. What I was doing at the time was a startup in the gimbal direction.

Jingxi: What level of technical capability did you achieve with the gimbal product at the time? Rian: At the time, gimbals only had two axes, and the footage looked very stiff when rotated by hand, which gave us the opportunity to innovate in control algorithms. We brought automatic panoramic stitching, night scenes, and stabilization—the most cutting-edge algorithms at the time—all to SOTA. We also used Kalman filtering for face tracking and connected the phone and gimbal via Bluetooth, making what was arguably the earliest "Camera Man standing on the desk."

Jingxi: How did the capital markets at the time view the Camera Man you were building? Was fundraising smooth? Rian: Back then, investors were all putting money into internet projects; nobody invested in hardware, and they thought hardware was a garbage business. But Professor Li introduced me and another co-founder to Wang Tao. Wang Tao thought our gimbal was very well made, and his product line really lacked someone who could lead a team, so he hoped we would join. But at the time we were fairly determined to start our own business, so we still chose to create our own brand.

Jingxi: So where did your money ultimately come from? Rian: As students, having 10,000 yuan in the account felt like a huge sum to us. In the end, we raised 1 million from friends and family around us. Crowdfunding brought in another 1 million-plus Hong Kong dollars, one of the highest crowdfunding amounts in this category at the time. At that point our product had an 80% gross margin, CPM of $5, shipments of hundreds of thousands of units, and we were quickly profitable. But at the time we were just a group of university students doing part-time entrepreneurship, and supply chain and management were all weak spots. After encountering manufacturers with stronger supply chain integration capabilities, we ultimately chose to sell the whole business.

Jingxi: You were very determined to start a business at the time—why were you willing to sell the company when the product momentum was so good? Rian: That's a very good question. I was just over 22 at the time and really wanted a result to prove myself, and the acquisition terms the other side offered were also very sincere. More importantly, we foresaw the endgame of this category: usually a hardware product has a six-month to one-year lead period, and once two companies in a track are doing the same thing, everyone starts competing on price. We were very afraid of others making the same thing, and the foreseeable future was becoming a battle of operations, which was not the result we wanted at the time. Rather than immediately entering a homogeneous commercial competition, we wanted to make more new things.

Jingxi: Then you went on to make a litter box—why such a big leap? Rian: I had just entered society at the time and had no occupational imprint. I felt that with engineering skills, a consumer perspective, and product aesthetics, I had unlimited possibilities. I had an underlying logic for choosing hardware categories: the demand has already been validated, and the competitive baseline is very low. At the time, the American smart litter box brand Litter Robot was selling well, but the product was unbearably hard to use. In our project at the time, we invented a front-and-back rotating structure that made the experience completely different, and crowdfunded $1.3 million, one of the largest crowdfunding projects in the industry at the time.

Jingxi: The litter box was later "paid tribute to" by many people in the industry—what lessons did you draw from your review? Rian: Looking back, in doing gimbals and litter boxes, we actually stumbled into "effective differentiation" without realizing it: what we did was directly related to revenue, and there were perceptual, non-parameter thresholds. Things with parameter thresholds ultimately get sold by the pound. We paid a big price in agriculture before we truly figured this out, and at one point it was very painful.

Jingxi: Why was the agricultural robotics experience so painful? Rian: In 2018 I started making fruit sorting robots. When we first started, we were one of the fastest-moving companies in the robotics category and the fastest in the industry to reach a scale of tens of millions, but this actually never broke through the level I reached in my first year selling gimbals. Looking back, the main reason was the difference between ToB and ToC business models. On one hand, making fruit sorting robots had no recurring revenue; on the other hand, we had to follow the growth cycle and harvest rhythm of fruit. The robot only worked for one month and then rested, and there was only one hardware iteration opportunity per year, with one trial and error requiring another year to validate. By the time we truly achieved a commercial closed loop, we found that what really drove the loop was not product strength and innovation, but grasping the laws of business itself—correctly and intelligently sharing profits with upstream and downstream. We could do this well, but it was not what we most wanted to do. However, it was precisely this experience that made us once again certain that we must pursue effective differentiation, and it also let our team go through a complete cycle together and develop a lot of tacit understanding.

Jingxi: Was it because agricultural robotics was so painful that you turned to AI hardware? Rian: Our investors also once questioned whether we had no positive feedback from making fruit sorting robots and wanted to escape it, but that was not the case. At the time we used AI for fruit quality sorting and were the team in the industry that understood fruit best. The turn came because of two things: on one hand, we were indeed better at ToC product definition; on the other hand, we saw a new opportunity in AI hardware. 02 Every time the underlying model gets stronger, Memory can do one more layer of things

Leiphone · Jingxi: How was the Flowtica team put together? Did they also come over together from the previous company? Rian: The team is now 50 people. The hardware core is comrades who fought all the way from gimbals to agriculture, a dozen or so people, partners who truly carried through cycles together. After moving from ToB to ToC, our requirement for talent density reached an unprecedented level. To make the best things, you need the most first-rate people with taste. And to attract such people, first of all your own standards must reach that level. We are also building the company as a product, making information flow within the organization as smooth as a product.

Jingxi: In the AI recording space, Plaud already exists on the market. Where does Flowtica's differentiation lie? Rian: It is already consensus that the boundaries of Agent capabilities are constantly expanding, but how to turn that into a product depends on what you care about. Some companies care about conversation—how to process dialogue better; we care more about Memory—whether AI can truly know this person, and then more proactively do things for him, give insights, and deliver personalized output. Actually, starting from the first-generation Scribe, we have been thinking about how to better understand users. FlowMark, which lets users mark key points during recording, is a product of this line of thinking. Marking is essentially the user telling the AI "this part is important to me"; it is Memory's earliest active input. This feature has now become an industry standard, but for us, from the very beginning it was never a tool feature, but the first step toward understanding the user. Every time the underlying model gets stronger, Memory can do one more layer of things. When the model is better at reasoning, it can extract deeper insights from your context; when the model is better at execution, it can complete more complex tasks for you. For us, this means new features can continuously be delivered from 0 to 1, and this is also the confidence behind our continuous expansion of the product line.

Leiphone · Jingxi: So you delivered another new product this year. Rian: Yes, this year we delivered a back clip with an e-ink screen. It is an agent-native piece of hardware: schedules, to-dos, emails, pre-meeting preparation—press the AI button to assign a task, and it will go do research, then deliver the results to you visually. We hope this hardware can be closer to users, ready to start at hand, like a companion. But for now we are not doing always on—not letting the device listen to you all the time, but keeping it ready so that it is within reach whenever you need it. The image shows Flowtica Verso.

Leiphone · Jingxi: Why must there be a screen? Can't a phone app do it? Or make it into earphones? Rian: An app absolutely can do it. But an e-ink screen sitting quietly there, visible at any time, can be reached visually instantly. As collaboration between people and AI becomes smoother and smoother, people's thinking will automatically become structured, and people need structured information; when receiving large blocks of information, listening is not very efficient. More importantly, intelligence now seems no longer scarce, and efficiency seems to have improved, but users still need to repeatedly explain the same context—every time they open an AI tool, they have to tell it from scratch who you are, what you are doing, and what you care about. This experience friction has never really been solved. We make hardware to let AI enter your context more directly, without you having to explain repeatedly. From the Scribe recording pen to the Verso e-ink card, and then to Apple Watch and Mac, we are building a multi-device ecosystem—in any scenario, there can be an entry point closest to you, letting AI seamlessly catch your context. Now big companies have also successively launched Memory features, and everyone is gradually feeling the power of "understanding you," while we have been doing this with a hardware ecosystem from day one.

Leiphone · Jingxi: When you made the second product, the first product had not been fully mass-produced on a large scale. How did you solve the marginal cost problem of the supply chain? Rian: For the first product, we prioritized components already validated by the market; stability came first. I come from a background in anti-shake cameras, so I understand the cost of aggressively choosing new chips. On a proven base, we use our own tuning to achieve the best results, such as the refresh strategy and model for the e-ink screen, and wait until volume increases before pushing for supply chain customization. Shenzhen's supply chain is mature enough; spend some money early, find good partners, and you will save it back later.

Leiphone · Jingxi: Your Scribe is priced at $159 to $209, and you also charge a subscription fee. What considerations are behind this? Rian: First, we have very high requirements for the quality of the hardware itself, whether in texture or design. Second, if hardware relies only on specs, it is destined to become increasingly competitive and have increasingly lower gross margins. We have confidence in charging subscriptions for two reasons: first, AI calls have real marginal costs; this is not charging out of thin air; second, Memory understands you better the more you use it, and the longer users use it, the higher its value. What we continuously deliver is not feature updates, but a personal intelligence that understands you better and better. The free tier is itself a complete product; the subscription unlocks deeper Memory capabilities.

Leiphone · Jingxi: What is the shipping rhythm and target going forward? Rian: Verso is now accepting pre-orders. The first-quarter target is tens of thousands of units, and then we will see whether we can push to 100,000 units per month. We do not want to trade price for scale; we want to trade experience for premium pricing.

Jingxi: Where will Flowtica go in the future? Rian: The underlying technology is still evolving rapidly. What we need to do is choose something whose underlying technology is still continuously evolving, and become first before it slows down. When everyone has their own Memory, the value of AI hardware will no longer be features, but how well it understands you. We will continue to explore hardware forms in the agent era, from recording pens and e-ink cards to more wearable hardware, expanding the product line. 03 New opportunities come from Memory, but the hardware rules of the old era still work

Jingxi: How did you deduce the new opportunity in AI hardware at the time? Rian: You can find a reference from a historical pattern. When AlexNet came out in 2012, convolutional neural networks brought a shock at the intelligence level, far beyond what traditional methods could compare with, but the training threshold was too high, and it was still a game for a few. The real change came in 2016—with the GTX 1080 paired with the maturity of open-source frameworks, you could train your own neural network with a consumer-grade graphics card once the environment was set up, moving from "scarce shock" to "mass accessibility." This pattern is now repeating. Large models and generative AI are following the same path as CNN back then—from high thresholds and high compute investment, gradually moving toward streamlining and popularization. Hardware intelligence is also improving exponentially, and eventually the phone in everyone's pocket will be able to run very good neural networks. Once the infrastructure matures, AI capability is no longer a barrier but a tool available to everyone. This pattern from scarcity to popularization happened with CNN in the past, and it is also happening on a new paradigm today, and within this, there is the opportunity for new consumer hardware.

Leiphone · Jingxi: Does the hardware company competition pattern you summarized back then still work today? Rian: It still works. A single big product can hardly support a good company; even if a certain product sells very well, it still faces market competition and needs to continuously lay out more product lines. What you fear is not competition, but the underlying technology being pushed to its limit. Take smart door locks as an example: after fingerprint, vein, and pupil recognition were pushed all the way, technological evolution has basically stopped, and all that remains is a price war. As builders, our nature is to constantly want to see new technology and constantly give users new things—for a category whose technology has already peaked, even if you reach first place, you will feel there is nothing left to do.

Leiphone · Jingxi: Then what can break this rule? Rian: When I was young, I wanted to change this rule; now I find I cannot change it, and can only choose a battlefield that belongs to me, which may be the only way to avoid it. The only path is to choose a direction whose underlying technology is still continuously evolving, and before its evolution slows down, become somebody in this industry and strive to be first. And the current explosion of AI technology excites me very much—every few months model capabilities take another step up, and new boundaries keep appearing. This feeling has never existed in previous tracks; you can see there is still a lot of room ahead to run.

Jingxi: You have made AI hardware your new battlefield, but there are also many opportunities in AI hardware. How did you converge onto the AI voice recorder track? Rian: Once intelligence is no longer scarce, what is scarce is Context—context, memory, understanding of a person. Last year, DeepSeek R1's chain of thought made AI better at thinking, and Qwen's function calling made AI better at doing things—intelligence capability and execution capability are leaping at the same time, and every step is validating this deduction. Meetings are the most natural first entry point for Context, so Flowtica's first product starts with an AI voice recorder, and its main use scenario is meetings. The image shows the Flowtica Scribe product.

Leiphone · Jingxi: Why insist on making it "anthropomorphic" rather than continuing to make tools? Rian: The endgame of tools will be eaten away by model capabilities. If you only do transcribe, only do summary, you will definitely be pushed aside—when ChatGPT first came out, Plaud connected transcription and summarization, which was enormously valuable; but if anyone today still only does this, they have no chance of winning. What is truly scarce is Memory. So from day one we made it anthropomorphic: voiceprint recognition can automatically identify speakers, and the Memory system accumulates the user's context. For example, at the end of last year, my year-end summary was written for me by Flowtica. It dug up details of my struggles that even I had forgotten, and finally told me, "Well done, keep it up." I was a little touched.