Summary

Vakee’s core bet on Bobby is that natural-language agents will replace complex financial apps—not merely add a chat entry point to RockFlow. Bobby is positioned as a “24/7 investment partner” spanning idea discovery, analysis and research, strategy generation, order execution, and portfolio management; the team plans to launch a Bobby-only product this year because in finance, where the combination of user needs is nearly infinite, “the future is simply agents plus databases.”
The product’s moat is not the chat interface, but the closed-loop capability built through 2 years of brokerage infrastructure and roughly 1.5 years of agent engineering. RockFlow was founded in 2021 and spent its first 2 years building a brokerage; it started developing “Trading GPT” in April 2023, launched it in May, already had an AI-native demo, and formally adopted an agent architecture in September. Vakee acknowledged that hallucinations were severe at the time, but API calls, document analysis, and incremental analysis were already usable. The guiding principle was: “Whatever we could do, we would do first.”
Early cases show both conversion potential and the non-generalizable randomness of returns. A UK user who had never bought a stock pre-authorized Bobby to buy Circle below $100; Bobby ultimately filled the order at roughly $80 and attached a stop-loss, after which the stock rose above $150. After Trump and Musk fell out publicly, an experienced trader bought Tesla puts through multiple rounds of dialogue; the trade gained more than 400% before being closed for profit. These are signs of product adoption, not proof of stable returns.
Bobby targets the simultaneous collapse of psychological and operational barriers, then keeps executing where general-purpose chatbots stop. Users can start with observations about Labubu, Ne Zha, war, or their work, receive explanations and risk analysis, and then confirm an order. When the host tested it with “I have $3,000—help me buy stocks. I want to make money,” Bobby built a 6-stock portfolio, explicitly naming SPY, QQQ, Apple, Nvidia, and Coca-Cola, and explained the logic behind each selection.
What truly separates the professional-user experience is long-term trading memory, personalized risk preferences, and composable orders—not a longer list of fixed features. Users have naturally taught Bobby to support “liquidate everything”; faced with a hypothetical $1M Tesla position, it would execute in batches to reduce market impact and slippage. The next or following release will proactively push ideas related to macro events, holdings, and watched themes, but it will not suddenly tell users, “You should buy this.”
Vakee’s investment method is not about uncovering mysterious information, but connecting partial knowledge from daily life and work to public markets. He cites the number and height of Taobao ad placements, his post-GPT-3.5 conviction around AI and Nvidia, and the growing piles of Pop Mart blind boxes in the office as signals that arrived earlier than financial statements. He bluntly says, “Fund managers don’t even have lives.” His entrepreneurial goal is to bridge ordinary people from “I noticed it” to taking investment action—not to predict which assets must rise.
Risk management is the non-negotiable premise of the “democratization of investing” narrative. Vakee says explicitly, “I can’t help you make money,” recommends that beginners start with $100, suggests considering put protection once a position is up double digits, or more than 10%, and repeatedly emphasizes that investing begins with “staying in the game.” Bobby’s promise is to add considerations, execution confirmation, and protective measures—not to transfer responsibility for the judgment to a model.

Deep dive

1. Bobby Is Taking Over the Entire Investment Loop, Not Just an Entry Point

Vakee describes Bobby as the first AI agent for financial trading and the user’s “24/7 investment partner”: it starts with daily idea discovery, moves into analysis and research, generates a trading strategy, executes the order, and continues managing the position rather than stopping at a research report.
His product analogy is “like having your own hedge-fund team,” combining traders, analysts, and risk managers. The point is not anthropomorphism, but having one natural-language interface orchestrate different professional functions.
Compared with the original RockFlow App, Bobby changes the unit of interaction: users once had to find pages and buttons; now they only need to express an objective. Vakee even imagines that “there may be no RockFlow App in the future,” with the main interface reduced to Bobby and an ongoing conversation.

2. “Agents Plus Databases” Will Replace Fixed Feature Menus in Complex Use Cases

Vakee’s criticism of traditional product management is that an app extracts the “greatest common denominator” from target users, then hard-codes those needs into features. Financial needs are highly individualized; once forced into pages, the product quickly becomes “10,000 features.”
Conditional orders are his central example. The team used to debate whether to build take-profit and stop-loss orders first or prioritize other order types, with each requiring front-end, back-end, algorithms, and interface work. But users may have 800 ways of placing an order. Bobby can recombine the underlying capabilities and generate “a conditional order that has never existed in this world.”
He uses booking flights for his parents to show that this is not unique to finance: there are 3 possible departure cities, layovers are preferably avoided, any layover cannot be too short, and arrival cannot be too late. The database has every field, but Ctrip’s fixed filters cannot express the full request in one pass.
Another example is asking a map to find non-spicy hotpot restaurants convenient by subway for 2 people living in Xierqi and Shuangjing. Vakee’s conclusion is categorical: “All of these relatively complex—slightly complex, even—scenarios will definitely become databases plus agents.”

3. Bobby Was Conceived 2 Years Before the Agent Boom

The timeline began in April 2023, when RockFlow started developing an “AI Strategy” page inside the app, originally called “Trading GPT.” It launched in May, attempting to connect “what happened” directly to personalized stock or options opportunities and push them to users in real time.
By May 2023, the team had already built an AI-native product demo and was exploring both bot and copilot forms. In retrospect, Vakee found the visuals less polished than today’s, while some interactions were actually more complete. The product was initially called Rockbot, not Bobby.
In September 2023, the team formally approved the project, with “Agent” written as the first item in the meeting notes. The decision came before products such as Manus were released: while “agent” was not yet an industry buzzword, the team had already committed to the architecture.

4. The Team Started with Hallucinations Rather Than Waiting for Foundation Models to Mature

Capabilities that were already reliable included calling large-model APIs, analyzing and digesting documents, performing incremental analysis, and orchestrating basic actions. The biggest flaw was severe hallucination. This produced a clear internal debate: start building now or wait for the underlying models to stabilize.
Vakee chose: “Whatever we could do, we would do first.” His boat analogy is that if a company operates at the application layer above foundation-model capabilities, upgrades at the bottom will naturally lift the product; waiting will not do the domain engineering for the team.
Much of the know-how in finance sits in product and engineering details. The earlier a team starts, the sooner it accumulates work that models cannot fill in automatically. He therefore viewed starting early as the optimal solution—not because the models were already good enough, but because domain-specific problems themselves require time.

5. The Moat for a Financial Agent Is the Hard Data and Trading Engineering

Vakee stresses that an agent product does not escape the engineering foundations of its domain simply because its interface becomes conversational. A classic financial problem is adjusted pricing and corporate actions: there may be hundreds or even thousands of possible scenarios, and each event must be handled correctly when it occurs. No vendor in the market can solve the full problem for a brokerage; IB, Futu, and RockFlow have all had to build it themselves.
The tolerance for data inaccuracies is extremely low, which explains the value of data products such as Bloomberg and Wind. Vakee calls this “the hard work”: it requires both an understanding of financial mathematics and a team that continuously handles edge cases in real operations rather than calling a generic API.
After RockFlow was founded in 2021, it spent 2 years building a brokerage, assembling a full execution and back-office trading system and serving users across different countries. Only afterward did the original team develop Bobby. Vakee’s view is that agents in verticals such as finance and healthcare still require product and engineering teams that genuinely understand the industry.

6. The Organization Had to Make Every Other Demand Yield to Bobby

Bobby began as the highest-priority project, then became an all-in effort over the past year, especially this year. RockFlow runs on biweekly planning and quarterly OKRs. For the past 2 quarters, it has explicitly halted every other app demand except trading stability: “all requirements yield to Bobby.”
The host argued that competition among large teams comes down to whether they dare to put the 10-plus people who can truly win on one project. Vakee said he was 99% certain about this future, with only 1% coming from the industry. He describes the team as 80%—90% INTJs and universally “very J”—people focused on the future they believe in, with almost no bandwidth to care what others are working on.

7. Productizing an Agent Requires Clearing the Speed and Cost Hurdles

Looking back on roughly 1.5 years of development, Vakee sees no major wrong turns but admits that early versions underestimated speed. Generating one result could take 10-plus to 20 seconds. “Slow is the original sin,” especially because young users will not wait; his acceptable threshold is 5 to 8 seconds, ideally 1 to 2 seconds.
The team subsequently rebuilt the agent orchestration system and its coordination methods to improve performance. Vakee’s productization requirements are: first close the loop and execute, then become fast enough; if the product is to be commercialized, token usage and total cost must also be controllable.
During internal testing, the team “shipped a new version every day” and iterated the model daily. Asked which metric he watched most closely, Vakee elaborated only on the first: engagement, especially whether users were willing to have multiple rounds of dialogue, because sustained conversation means Bobby is genuinely resolving a question or action problem.

8. The Circle Case Shows How a Beginner Crosses the First-Trade Barrier

A UK user had been registered with RockFlow for some time but had never bought a stock. On Circle’s listing day, he asked Bobby to buy a specified amount after trading began, then went off to handle other things. It was the first stock he had ever owned.
Bobby did not place an unauthorized bet: it first asked about the price and proposed executing below $100. Only after the user explicitly confirmed did it place the order. The final fill was around $80, with a stop-loss attached; the price later rose above $150, and the user had not sold.
The user shared the full conversation and screenshots of the gains in a group, saying that without Bobby, his first trade might have been delayed indefinitely. When the host asked who would be responsible for a loss, Vakee reiterated: “Every order has been confirmed.” The amount and price, or price range, must first be authorized by the user.

9. The Tesla Put Case Shows That Experienced Traders Need a Personalized Agent Too

When Trump and Musk were publicly feuding, a user who had traded Tesla for years asked how to respond. After multiple rounds of discussion, Bobby suggested considering puts and helped him choose a specific option. The user ultimately took profit after the trade gained more than 400%, becoming the day’s “top earner.”
The user later said he could simply tip Bobby 1% or 5% of the gains. Vakee did not attribute the entire return to the model, adding an important boundary: “You made the final decision yourself.”
More important was the user’s first experience of an agent understanding his “past trading behavior and preferences”: how he trades Tesla, what structures he prefers, and how he manages risk. His reaction was, “So this is how you can trade,” and he suggested making the agent fit his history like an investment assistant.
This contrasts with the Circle beginner. Experienced traders might previously have found RockFlow too simple, while beginners still found brokerages difficult to operate. A natural-language agent no longer compromises through one fixed interface; both ends can express their full needs in their own way.

10. Natural Language Removes Both the Psychological and Operational Barriers

Vakee divides the investment barrier into 2 layers: first, the psychological barrier of “thinking it is too hard”; second, the operational barrier of not knowing where to find a feature or how to place an order. Voice or text lets users describe observations from daily life as if talking to a friend, without first mastering financial terminology.
Users can ask what the Israel war, a Trump statement, the Labubu craze, or the popularity of Ne Zha might mean for different companies. The host pushed back that ChatGPT, Yuanbao, and Doubao can answer those questions too. Vakee acknowledged that at the level of “putting an idea into words,” their interaction value is the same.
The difference comes with “what next?” A general chatbot explains and stops; Bobby asks whether the user wants to execute, how much to execute, and completes the order after confirmation. Vakee says the point of a vertical agent is to “help the user get the thing done.” Saying it is not enough.
The host tested it with the most basic instruction: “I have $3,000 right now. Help me buy stocks. I want to make money.” Bobby built a 6-stock “hexagonal portfolio,” explicitly naming SPY, QQQ, Apple, Nvidia, and Coca-Cola, and explained the allocation logic for each.

11. Proactive Alerts Will Increase, but Trading Authorization Still Has a Boundary

The team had never prioritized a “one-click closeout” feature, yet users simply told Bobby, “Sell everything for me.” The agent called the existing underlying capability and completed the request. Faced with a hypothetical $1M Tesla position, it would place the orders in batches to smooth market impact and slippage.
The next or following release will proactively provide ideas tied to macro events such as wars, news related to the user’s holdings, and changes in watched themes such as Circle and Coinbase. But Vakee draws a line: Bobby will not suddenly appear and command, “You should buy this.” A reminder is not a decision instruction.
The host said that one friend working at a Hong Kong family office told him one of the clients’ most pressing needs was for the team to call and provide reminders. That 4- or 5-person team had historically relied on phone calls to handle the vague research questions and explicit order instructions of clients in their 90s.
Vakee describes Bobby as a master agent supported by countless smaller agents. Unlike hard-coded RPA, it continues learning from each use case, user feedback, and model upgrades. He believes this iteration speed will quickly push its capabilities beyond those of many traditional financial roles.

12. Open Agents Suit High-Tolerance Industries; Critical Verticals Must Hold the Workflow Floor

Vakee believes entrepreneurs must first choose between a general-purpose and vertical architecture. For tasks such as writing emails, planning trips, or generating documents, users might manage a score of 70 themselves, and an agent scoring 60 to 65 may still be acceptable because a mediocre result is usually not fatal.
Finance, healthcare, and precision manufacturing are different: “A score of 70 and no 70 at all could be fatal.” When accuracy or completeness must clear a hard floor, a workflow-based structure makes it easier to keep results consistently above 70.
RockFlow chose a vertical workflow from the start, so even as models and the technology stack changed continuously over more than a year, the overall architecture did not need a major redesign. Vakee also denies that there is 1 or 2 token-saving secrets; the answer is to keep applying the best engineering practices available at the time around clearly defined use cases.

13. Buying Jidong Cement at 9 Made Investing Feel Familiar from the Start

The first stock Vakee bought, at age 9, was Jidong Cement. In his childhood, market quotes were broadcast on television from 1 p.m. to 3 p.m., newspapers printed candlestick charts, orders were placed by phone, and brokerages had rooms for major clients. Adults around him all traded stocks, unconsciously lowering both the psychological and knowledge barriers.
He later began trading US stocks for the same environmental reason: after returning to China and joining Baidu, his internet colleagues were early participants in Chinese ADRs and US equities. Vakee uses this to explain his product belief: whether people invest depends heavily on whether they have information, language, and behavioral examples around them.
His undergraduate studies spanned advertising and finance, alongside calculus, probability and statistics, programming, and psychology; he later moved into data science for graduate school. Vakee calls this “generalist logic,” citing classmates with backgrounds in law and philosophy who were nevertheless the strongest coders as evidence that foundational abilities often transfer across disciplines.

14. From London Quant to Baidu’s Phoenix Nest, Every Step Built Bobby’s Required Capabilities

Around 2012-2013, Vakee was working in quantitative finance in London when he received offers from Singapore’s sovereign wealth fund GIC and Baidu. He firmly chose to return to China for mobile internet because the wave depended on population dividends; the opportunity was more likely to emerge in China or the US than in the UK or Singapore.
At Baidu’s Phoenix Nest, he was already a user of AI, although the common terms at the time were machine learning and neural networks. The Phoenix Nest team numbered only dozens of people and spent 1 to 2 years getting revenue above the PC business, giving him a firsthand view of AI connecting to measurable commercial outcomes.
He then joined Baidu’s investment division to cover technology investments, followed by VC, private-market, and public-market trading. Moving from “building AI” to “investing in AI” and then founding a trading platform, he sees the nearly 10 years before entrepreneurship as a period in which “every step counted.”

15. Baidu Saw AI; Its Business Choices and Organizational DNA Were Misaligned

Vakee begins with praise: without Baidu’s Phoenix Nest and investment division, he would not have had his later opportunities. In 2012 and 2013, Baidu spent heavily to build laboratories for deep learning, computer vision, and big data, attracting talent in Beijing and the US. That investment led most Chinese internet companies by years.
The host’s follow-up was blunt: there had been “a lot of groundwork and praise,” so how exactly did the regret arise? Vakee answered that Baidu was an engineering culture obsessed with efficiency, skilled at optimizing performance through search, Phoenix Nest, and algorithmic tuning. Advertising and recommendation were naturally the AI use cases best suited to it.
He believes the problem predated large models. Baidu poured substantial effort into O2O businesses such as Nuomi and Baidu Waimai without getting the organization to accept the lower efficiency, different talent structure, and cultural gap that offline operations inevitably brought. The AI team was not behind; the company chose “things that did not fit its genes.”
Not long after joining Baidu, Vakee suggested to Robin Li in his first one-on-one that Baidu should build a financial business and create a Bloomberg. Ten years later, he would still give the same answer. The founder-product fit lesson is that founders must genuinely believe in the target before they can tolerate years of investment, after which organizational execution turns that belief into reality.

16. GameStop Turned “Value Investing” into a Generational Expression of Values

Retail traders’ battle with Wall Street was RockFlow’s trigger, not the complete answer. Vakee saw young people using investment not only to make money but to express an attitude: “If I believe it has value, that is value investing.” They had childhood memories of GameStop, so Wall Street should not be allowed to short it.
For this group, investing became part of a lifestyle. Further research showed that young people around the world wanted to buy companies such as Tesla and Apple, but outside China and the US, there was little supply of usable US-stock brokerages and market penetration was low.
Vakee uses 4 criteria to screen entrepreneurial directions: the industry must be large enough; it must naturally suit AI and data iteration; he must have a “passion that makes it impossible not to do”; and the team must genuinely be capable of execution. He has invested in nearly 30 startups and sees one common reality: “Nothing is easy for anyone.”

17. The Company Was Designed from Day One as All-in-One and AI-Native

Vakee defines the new species along 2 axes. All-in-one means one application can trade US, Hong Kong, Vietnamese, and Japanese stocks, as well as futures, options, FX, crypto, lotteries, event contracts, and other asset classes, rather than fragmenting users across markets or categories.
The second axis is AI-native: not embedding AI into an old brokerage, but redesigning the experience of discovery, research, strategy, and trading. Existing apps do not match Gen Z users’ demand for simple, intelligent experiences, creating room for an incremental market.
The corresponding team threshold also has 4 parts: understanding AI; understanding financial mathematics, options pricing, the Greeks, trading systems, and clearing and settlement; understanding compliance; and having global operating capability. Vakee emphasizes that “knowing how to trade stocks” is not the same as understanding financial engineering.
That is also why he was willing to build a difficult To C product. People often ask why he would start a company when his personal account has consistently made money. His answer is a sense of mission: “Investing should not be a difficult thing.” RockFlow may be the work toward which his experiences collectively pointed.

18. Ordinary People’s Daily Observations Can Be Alpha Professional Capital Does Not Have

Vakee’s own investment information has mostly come from work rather than from searching outside it. At Phoenix Nest, he could open Taobao and look at the number and height of the first few ads to roughly judge whether results would meet target. Later, when investing in AI and buying Nvidia in 2015, he translated professional knowledge into a public-market position.
His distinction is that life and work provide the probability edge, while the choice of stock versus options, position size, and protection determine how that edge is leveraged. After GPT-3.5 launched, many people already believed AI was the future, but they did not take the next step of identifying the platform winner or connecting that insight to Nvidia stock.
The host added his own missed opportunity: a year earlier, he had heard that Labubu was rapidly surpassing Molly, but failed to turn that into a Pop Mart trade. All he could say was, “Money was falling from the sky, and you didn’t reach out to catch it.” Similarly, some people use Pinduoduo or drive a Tesla for years without ever owning the corresponding stock.
Vakee’s sharpest judgment is that “fund managers don’t even have lives,” so they may simultaneously miss Pop Mart, Laopu Gold, and Mixue Bingcheng. DAU is simply more people around you using something; growth may also be your own continued consumption. Ordinary people’s edge is seeing the parts of life that professional institutions have not entered.

19. Investing Is Democratized Only When It Is Tied to Risk Management

At the end of 2023, Vakee bought a large position in Robinhood at around $8 and treated it as a self-test of an entrepreneur’s judgment: if you build a trading platform but fail to see an opportunity in the same sector, your industry judgment is flawed; if you see it but do not act, you cannot demand that the team fix a bug within 3 hours.
He wants consumers to become shareholders too. When Square listed, he helped many customers subscribe to its IPO, while Robinhood users owned shares in the company. He also once suggested that Bilibili give stock to premium members. By contrast, Didi drivers contributed to the company’s growth but received only labor income and no capital return.
The host’s final push was that ordinary people may see the signal but still need rational checks. Vakee responded that Bobby is not a “voice function for placing orders.” The same Labubu observation could produce different answers a year ago and today, when the stock is already at a high level, because Bobby will explain the stock’s position, how long it has stayed there, and the risks.
The boundary is repeated again and again: “Bobby has never said it can help me make money.” Vakee recommends starting with $100, considering puts once a position is up double digits, or more than 10%, and avoiding an easy all-in. The first priority in investing is to “stay in the game, then wait for the opportunity that belongs to you.”

Talk with 王鹤: Embodied AI’s Fringe History and Capital-Fueled Disorder Macro Conversations 88: Hong Kong, U.S., A-Shares and AI Investing