100 AI Creators is a weekly series featuring conversations with China’s leading minds in artificial intelligence. As technology evolves, their perspectives shed light on the ideas driving the AI era across borders.
Before building Bobby, Vakee Lai had already been trading stocks for over 20 years.
She started investing at the age of 9—three years later than Warren Buffett—and could make trading decisions based on instinct. After graduating from Imperial College London, Lai joined a quantitative fund, where she used machine learning to trade futures and derivatives. In her twenties, she made her first RMB 10 million (USD 1.4 million) shorting a US stock.
Her account balance inflated like gaming points. To Lai, making money was easy enough that it began to (almost) feel boring. She had little interest in indulging in extravagances such as handbags, travel, or luxury goods. Most of her shopping is done on Pinduoduo, and she never flies business class unless it’s somehow faster.
When trading starts to feel as simple as breathing, Lai thought she needed to find a wilder game.
After working in product development at a major tech company and spending five years in venture capital, Lai launched her own startup. Her first product was RockFlow, a US stock brokerage platform designed for global users.
RockFlow started off using AI to simplify the investment process, offering features like daily trading recommendations, automated copy trading, and streamlined options trading. It sought to turn a complex financial platform into something that felt more like a game.
But the product, which launched in 2021, was still just a “smarter tool.” It could assist with trading, but it didn’t understand why users wanted to trade, much less execute them on their behalf.
That’s where Lai’s new project, Bobby—an AI agent—comes in.
Trading in natural language
Bobby’s goal is to complete the entire investment workflow—from intent parsing to strategy generation to order execution—using just natural language. For everyday users, it means simply expressing what they want in plain speech. Bobby does the rest.
For example, when you’re browsing Pop Mart’s website, Bobby might pop up with a message:
“You’ve spent RMB 2,000 (USD 280) on Pop Mart this month. Labubu is trending on TikTok. The secondary market price for the Labubu x Vans collaboration has a premium rate of 1,284%. Should we increase your exposure to the collectible toy market?”
This marks a leap from chatbots like BloombergGPT or Morgan Stanley GPT. Those merely answer questions, resembling manual tools like the earlier version of RockFlow.
To Lai, AI agents aren’t upgraded apps, but a new generation of user interfaces (UI).
“All apps will disappear in the future. They will all be replaced by [AI] agents,” she added.
Take a more extreme example. Suppose Bobby is told that “I hate Donald Trump,” the agent might interpret this attitude into a portfolio consideration and suggest avoiding or shorting Trump-related stocks, while recommending long positions in fields like renewable energy, given Trump’s energy policies. Bobby would then ask about your investment goals, helping to adjust investment strategies based on risk appetite and expected returns.
In this world, a person’s values and emotions can become executable trading signals. That was the starting point for Lai’s Bobby.
When trading is a lifestyle
To Lai, trading is as natural as eating is to foodies, or flying is to travel buffs.
For most people, investing feels complicated and intimidating. But for her, every day is filled with trading opportunities, ready to be seized anytime, anywhere.
For example, during this year’s Lunar New Year, DeepSeek became so viral that even her relatives back home started chatting with AI. The RockFlow team quickly applied for access to DeepSeek-related cloud services. She promptly bought shares in Alibaba, convinced that its cloud computing business would benefit from the surge in compute demand driven by open-source AI models.
On a trip to Japan one year, she noticed that Sagami’s ultra-thin condoms were sold out at every convenience store she visited. She researched the company and found its exploration of ultra-thin polyurethane materials to be more advanced and aggressive than Okamoto’s. It was also a bestseller on cross-border e-commerce platforms. That stock ended up bringing her fourfold returns that year.
Trading, for Lai, is a way of life, and she firmly believes in it.
Rather than indulge herself materially, she’s passionate about challenges that test her discipline and limits. She once passed China’s bar exam in just nine days, and earned the ACE personal trainer certification with the highest score.
Lai later channeled that drive into AI. Over the past decade, she has either been developing AI products, investing in AI, or launching AI-related ventures.
AI in the future of trading
When AI Now! met with Lai, the US had just fired the opening shot in a fresh tariff war.
The cafe—where the interview was held—buzzed with anxiety as markets jittered across everyone’s phone screen.
Lai was asked: “What can Bobby do now?”
“Bobby, find companies least affected by tariffs and those with potential for exceeding growth expectations in the next three months,” Lai said.
Seconds later, Bobby replied:
“Consider sectors with tariff exemptions, like AI infrastructure, data centers, and semiconductor manufacturing equipment. These are areas that might receive tariff exemptions or delayed implementations. Major US-based manufacturers with domestic production facilities may also outperform, as they avoid import tariffs.”
“Additionally, consider configuring top companies in rigid-demand and defensive industries. Please refer to the tariff opportunity stock list,” it added.
Then came the kicker: “Would you like me to automatically place orders if these companies’ share prices drop by more than 10% within a week?”
In the AI era, some people use large models to make videos and presentations. Others flirt with DeepSeek’s chatbot. Some want an eternal AI-powered pet cat.
But Lai believes the future of trading will soon become a battle of algorithms. And everyone will need their own Bobby.
The following transcript has been edited and consolidated for brevity and clarity.
AI has no emotions, and that’s why it makes money
AI Now! (AN): Many financial products today have integrated large language models (LLMs), with examples like Bloomberg’s BloombergGPT and Morgan Stanley’s upcoming copilot. What makes Bobby different? Is it just a chatbot that can place trades?
Vakee Lai (VL): If Bobby were just a chat-enabled trading tool, I wouldn’t have bothered building it. Have you ever asked one of those financial chatbots “what to buy right now?”
AN: They analyze the market, provide some suggestions, and inevitably follow up with the usual disclaimer that “investing carries risks.”
VL: Exactly, because they only respond, they don’t decide.
Bobby is different. If you say, “I hate Donald Trump,” it won’t just summarize the impact of Trump’s policies on the market. It’ll ask directly if you’d like to remove Republican-affiliated stocks, or highlight that the renewable energy sector might benefit.
AN: So it’s a more advanced form of semantic understanding?
VL: No, it’s a completely different logic. Chatbots wait for your questions and respond. Bobby anticipates. If it senses that you’ve been reading semiconductor news but don’t own any chip-related stocks, it might ask if you’d like to monitor TSMC’s earnings, and whether you’d consider buying in if the stock exceeds expectations.
You can also teach Bobby your own analysis logic.
AN: It sounds like Bobby is solving the issue of turning trading intent into execution, which is something users may struggle with?
VL: It’s hard for most people. Take the recent US stock market slump as an example. Before tariffs were imposed, everyone knew there was a high probability of a major drop, but deciding when to act, how much to sell, and what to sell requires constant market research and timely execution. Most people have the intent but not the action, either due to inertia or simply not knowing how to proceed.
Even basic strategies like “buy low, sell high” are hard for most people. When their stocks rise 30%, they are reluctant to sell, hoping it’ll go up to 50%. But when they don’t take profit, a big drop could follow and at that point, panic selling usually kicks in.
But AI doesn’t do that, because it has no emotions.
AN: Are you saying emotions make it impossible to win in the stock market?
VL: Emotions bring greed, anger, and delusion. Trading requires you to go against human nature. AI doesn’t have that baggage, and that makes it better at making money.
AN: What if I asked Bobby to adjust my portfolio so I have zero risk and 100% return?
VL: See, that’s greed. Bobby would honestly tell you that it’s impossible and then help you turn your get-rich-quick idea into a realistic strategy with, say, a 20% annualized return.
AN: But won’t the trust barrier of AI trading agents be too high? If an agent botches a PowerPoint, you can just fix it. But if it makes a trading mistake, it could lose a lot of money. Won’t users worry?
VL: We’ve built in safewords. You can say things like “Bobby, just monitor but don’t trade,” or “Call me if Nvidia drops to USD 98,” or “Limit each order to no more than USD 10,000.”
In fact, most users are more reassured after trying it out, precisely because AI doesn’t suffer from human weaknesses.
About starting a company
AN: Your first venture was RockFlow, a Gen Z-oriented brokerage platform. From there, did you feel that building Bobby was a necessity, or was it because everyone started diving into AI agents?
VL: From day one, My vision for RockFlow was to make investing simpler. Over the past two years, we tried many ways to realize this vision, like radically simplifying the app interface. But I felt it wasn’t enough.
By 2025, I’m convinced that the new generation of young users—especially those who grew up with AI—can fully operate through natural language for transactions. All you have to do is speak, and Bobby will quickly transform your thoughts into trading strategies and execute the trades for you.
AN: Plenty of brokerage platforms are integrating general-purpose large models to provide investment advice. Why didn’t you just add a DeepSeek-powered chatbot to RockFlow? That seems like a natural upgrade.
VL: General-purpose large models can indeed provide standard analysis, but it’s far from enough for real-world investing. When you need to analyze and execute based on individual users’ needs, factoring in their trading preferences with real-time market analysis, such models tend to fall short. They can transmit bulk information to users, but can’t complete the decision-making loop.
Bobby uses a hybrid of workflow logic and LLMs or agent models, maximizing creativity that AI can offer while keeping costs and risks under control.
The key is that we run all incoming data—user behavior, market trends, public sentiment—through a second layer of financial expertise before Bobby sees it. That’s the only way to ensure it really understands what users mean and can execute accordingly.
AN: Some general-purpose agents now integrate vertical knowledge bases. Why not wait for those to get better, instead of building a vertical agent like Bobby?
VL: That comes back to a fundamental issue. There are two types of user needs in the world:
- Life-or-death needs, like financial trading or medical diagnostics. Most people can’t even hit 70% proficiency here, and falling short means serious consequences. But getting to 70% is very much achievable through defined paths.
- Nice-to-have needs, like making PowerPoint slides or planning a vacation. Falling short doesn’t carry the same risk.
General agents are fine for the second. But for the first, you need vertical agents because:
- The data is different: We need millisecond-level trading data and real-time portfolios.
- The responsibility is different: A bad stock tip can wipe out someone’s saving.
- The decision process is different: You can’t just offer “possibilities”—you need executable judgments.
AN: Like you wouldn’t trust a general practitioner to do heart surgery?
VL: Exactly. Bobby was trained from day one like a finance major. Every decision it makes reflects that.
AN: Some say workflows can only handle narrow, retrieval-based tasks. How do you make Bobby actually “understand” finance?
VL: A workflow is just a tool. The real difference is in how you use it.
Industry knowledge is the foundation. The real leap is this: we’ve made workflows that can dynamically generate context-specific trading insights—from a user’s risk profile and positions to real-time intent and sentiment—combined with deep domain knowledge from our team.
It’s not about fetching facts, but rather having a trained trader who can respond instantly to market signals. In scenarios tested on RockFlow, Bobby is orders of magnitude faster than a general LLM.
We’re building toward a world model that learns and adapts to market changes in real time: true intelligent, high-efficiency financial decision-making.
Understanding users and building proprietary systems
AN: And what about understanding the user? How does Bobby know who you are and what you want to do?
VL: Bobby is integrated with RockFlow’s order execution system, live market feeds, and user databases. Think of it as a hedge fund manager that works for you round the clock.
Say a user tells Bobby, “I just lost my job. I want a safer portfolio.” It will lower the risk profile, suggest government bonds and high-dividend stocks, and set dynamic stop-loss lines. When markets fluctuate, Bobby might proactively ask, “The Federal Reserve may raise rates. Should we adjust your bond allocation?”
AN: If vertical agents represent the future, why haven’t we seen platforms like Robinhood or Futu move in this direction?
VL: Building an AI-native investing experience requires rethinking your entire tech stack. Legacy brokerages are too tied to their existing infrastructure. The more successful these traditional players are, the harder it is for them to abandon their existing infrastructure and user experience.
Previous-gen brokerages focused on better mobile UI versus desktop brokers like Interactive Brokers. That made sense for Gen X and older millennials. But today’s Gen Z users want a whole new way to invest.
We designed RockFlow from day one for AI-native users, building our own order execution system and AI infrastructure. Nobody else is doing that globally.
AN: Wait, building your own order system? What does that actually mean?
VL: For a brokerage, the order execution system is like TikTok’s recommendation algorithm. If you use someone else’s system, it’s a black box. You can’t train your model on it. We had to build our own so we could access structured, continuous user behavior data. This is what lets us understand users’ decision paths and optimize over time.
From the beginning, we designed RockFlow’s system with machine learning in mind. This data isn’t static. It is raw material for training each user’s personalized Bobby.
AI agents as the final evolution
AN: It sounds like you’ve been preparing for AI agents since day one.
VL: I dug up the meeting notes from our first discussion about Bobby—it was in September 2023. From the beginning of RockFlow, we designed everything—trading systems, data flows, product architecture—with an AI-native goal. But our thinking around agent architecture became clearer over time. We made some wrong turns too, and those were valuable lessons.
Every generation of user-facing products has its mission. In the AI era, your product’s mission can’t just be to add more features or make the UI cleaner. It’s no longer about adding or removing a button.
AN: So what is the mission of products in the AI era?
VL: For the first time, products can actually understand users and serve them proactively. A great mobile internet product helps users operate more efficiently. A great AI product means users don’t need to operate at all. They are simply served.
So where the app once guided you through steps like how to buy options or how to find the right option, today AI can directly understand your intent: “I want a 20% return,” or “I want to avoid this market crash.” That’s a fundamental paradigm shift.
That’s why I say agents aren’t upgraded apps. They are the next-generation UI.
AN: What is Bobby’s mission?
VL: My original motivation was that investing is deeply personal. It reflects a person’s worldview and values.
Like we talked about earlier, many people have views and reactions to daily events, but get tripped up by tactical details. They can’t act on their insights. People often say trading is “turning cognition into profit,” but for most, the hardest part is going from knowing to doing. Bobby exists to bridge that gap.
AN: So whether I love or hate Donald Trump, I can trade on it?
VL: Absolutely. Both can be investment signals. Both can make money.
AN: So it’s no longer about using a tool, but about being served by an agent?
VL: Right. From that perspective, all apps will eventually disappear, and agents will take their place.
AN: I believe in that future too. But aren’t you worried you’re too early? That you’ll end up a pioneer who gets arrows in the back?
VL: I don’t think about it that way. For the past ten years, I’ve either invested in AI or built AI. My path and experience all led to this moment. I was meant to build Bobby. The brave get to enjoy the world first.
Quick takes
AN: What’s the most jaw-dropping moment AI has given you in 2025 so far?
VL: When DeepSeek displayed its deep reasoning visually for everyone to see.
AN: What’s something in AI you were skeptical about last year but have completely changed your mind on?
VL: Text-to-image generation. Early on, it was incredibly hard to generate images that were accurate, especially with consistent character poses or legible text. I thought commercialization was far off. But starting with diffusion transformers, new tech and products blew past expectations. Now it’s usable across real-world scenarios.
AN: If you could shut down one AI product or trend you think is completely misguided, what would it be?
VL: Any product that’s just layering generic AI features on top of OpenAI, or those that haven’t built a true closed-loop application. They are too easy to be wiped out by OpenAI’s next update, like when GPT-4o could instantly generate Studio Ghibli-style animation and killed off a bunch of startups. A strong business moat matters.
That’s why, for me, the question isn’t whether AI is being used, but rather what problems are being solved. In vertical scenarios, the real challenge is abstracting core user needs and building long-term value. This space still needs better AI product managers.
And as for Bobby, we didn’t build it to show off AI. We built it to make investing easier. If one day it fails to do that, we’ll shut it down. No need to get sentimental.
AN: What’s being underestimated in AI right now? What’s being overhyped?
VL: Demand for computing power is seriously underestimated. Artificial general intelligence (AGI) is overhyped. People think throwing more GPUs and training larger models will bring AGI closer. But the real bottleneck is in applications. I believe the next few years will be the breakout era for vertical agents. And that’s going to be a long game. Each use case needs customized compute optimization. Just like when electric vehicles took off, but charging infrastructure became the real constraint.
AN: What are you most looking forward to in 2025?
VL: A “Cambrian explosion” of vertical applications. I expect to see agents that truly reinvent user experiences in areas like finance, healthcare, education, travel, and supply chain—not just glorified chatbots.
Imagine a travel agent that plans personalized itineraries, negotiates pricing, and pays for you. Or an education agent that adapts learning paths to your pace and strengths. None of this requires AGI. Current tech plus vertical data is enough.
AN: And finally, three books you’d recommend?
VL: “The Pleasure of Finding Things Out” by Richard Feynman, “What Makes You Not a Buddhist” by Khyentse Norbu, and “Deng Xiaoping and the Transformation of China” by Ezra Vogel.
100 AI Creators is a collaborative project between AI Now! and KrASIA, highlighting trailblazers in AI. Know an AI talent we should feature? Reach out to us.