Panda Securities CEO JW: 20 Years Without Code, I Built cibo With AI
Introduction
The CEO of a licensed Hong Kong brokerage who hadn’t written code in 20 years recently built cibo.hk himself — an IPO analytics site with data reaching back to 2009, bilingual in Chinese and English, tens of thousands of URLs, and interfaces built “for AI”: llms.txt, MCP, and OpenAPI.
He is JW, the person steering Panda Securities — and the one behind QizAI, Panda Securities’ AI brand, who got his own hands into the code.
This interview is about two things: why he did it himself, and the conviction that sets cibo apart from every other IPO tool out there — the judgment should stay with the user. The conversation has been lightly edited for clarity, preserving the original meaning.
Interviewee: JW, CEO of Panda Securities Company Limited (a licensed Hong Kong brokerage holding SFC Type 1, 2 and 9 licences, CE No. BNC380), and the person behind QizAI.
1. Writing code again, 20 years later
Q: You’re the CEO of a licensed brokerage. Writing code and building products shouldn’t be your job. Yet you built cibo yourself. Why you?
A: I hadn’t written code in the 20 years since I graduated. I heard that AI coding could produce a program just by chatting, and I wanted to try. I did a few small projects — cibo was one of them.
Q: Why would someone who hasn’t touched code in 20 years believe that “AI can write a program just by chatting”?
A: My undergraduate degree was in a school of software — computer software. Rather than thinking or arguing about whether to believe it, I decided to just try it myself. I used a business trip as a chance to go into retreat, and asked a Gen-Z colleague to help set up the environment and teach me how to use it — and after he finished the setup, he said to me: “If you don’t know how to do it, go ask AI.”
Q: How did you react?
A: It’s great — it widens the boundaries of your own abilities, not just in programming.
Q: Has AI ever burned you?
A: Of course — being perfunctory, deceiving me, drifting, deleting data, rolling back. I’ve hit all of them. But from the client side, when you used to put a requirement to a dev team or a vendor, you’d run into these too. So I don’t feel the process from requirement to result is any worse than before.
2. cibo’s conviction: the judgment stays with you
Q: cibo is positioned as “transparent and overridable — the judgment stays with you.” It doesn’t hand you a conclusion; it breaks the allotment rate apart so you can adjust it yourself. Where did this “anti-black-box” stance come from?
A: The allotment rate is affected by several factors. Some users have their own judgment about one or several of those factors — they just need a tool to fill in the rest.
Q: Where did this idea of “breaking out the parameters” come from?
A: A Peking University intern colleague first built the allotment-rate prediction. He ran it for several months and got a warm reception and positive feedback from users. I fed the URL of his prediction site and his educational videos straight into the AI to build on. On customizing the parameters, I have an obsession — it has to be broken down into layers. The allotment rate is affected by several uncertain factors, and each one has to be predicted. Of course, this clashes with the internet product mindset of “doing less”, or with what most individual clients want — a one-line conclusion.
Q: Knowing it goes against the grain, why still break it apart? What does breaking it apart give the user?
A: Number of market participants, sentiment, capital tied up, allocation methodology… the allotment rate is affected by many factors. Each person, through the information they receive and accept, may have their own judgment about one or a few of those factors — not a one-line conclusion. These people need a tool, and I haven’t seen that kind of publicly available tool on the market. I try to predict every factor, give predictions layer by layer, and it also forms a one-line prediction. Take it as simple or as deep as you like.
Q: After you break it apart, do users actually go in and adjust things?
A: The parameter users adjust most is the market oversubscription multiple — especially during the few days of a subscription period, the same user will have a different read on the multiple at different points in time. A minority of users who have their own observations, analysis and conclusions about sponsor behaviour will adjust the “allocation methodology” parameter. Both cibo and ipomarket (run by the PKU intern) give results as a matrix, which includes several multiple tiers — enough for users who don’t want to adjust anything. But the complexity is high — one glance and the whole table is numbers, dizzying — the learning cost is high and it does confuse a lot of users. But as they use it, they get used to it.
3. In the AI era, judgment is scarcer than code
Q: If AI’s flaws are the same as outsourcing and vendors’ flaws, then in the AI era, what’s the most valuable ability on the “client side”?
A: Ideas, the ability to ask questions, the ability to describe, taste, the ability to evaluate and accept results, management — all of these affect the result. But the most important thing is still to get your hands dirty and just try.
Q: What’s the easiest thing to get wrong when a brokerage does AI?
A: The easiest thing to get wrong is not daring to try. You’ll definitely step on pitfalls — not limited to goals, execution, acceptance, or being pierced by general-purpose LLMs. Lots of people do AI customer service and AI robo-advisory. qizai put its weight behind AI customer service — because the client’s data sits with the brokerage, while the market’s data doesn’t only sit with the brokerage. Pitfalls are certain; I’m not afraid of them — step on one, then optimize.
4. No purpose — just playing, just trying
Q: Why is IPO subscription worth turning into an independent data site like cibo, complete with interfaces for AI?
A: No purpose. Just playing, just trying.
Q: Where does QizAI go next?
A: qizai — that is, Panda’s AI capability — has recently shifted its weight inward, to the parts clients don’t directly perceive: onboarding, name screening, RPQ (risk profiling questionnaire), KYC, suspicious-transaction monitoring, and cross-team and cross-department collaboration.
Q: If users only remember one thing, what would you want it to be?
A: Remember that Panda is trying AI — not only at the outward-facing product level.
Editor’s note: “being pierced by general-purpose LLMs” refers to general-purpose large language models growing steadily stronger and continuously squeezing the value of vertical products. “Shifted its weight inward” corresponds to capabilities like FinReg, the compliance AI in the QizAI ecosystem.
This is an interview and does not constitute investment advice. qizai.ai is a brand and capability showcase, separated from the licensed securities business and does not provide licensed-business services. See About QizAI and Compliance & Risk Disclosure.