AI Trading in China's Stock Market — AgentBull Is Betting Real Money on It

Richard and his team at AgentBull (煜马) built an AI-native quantitative system using OSINT data and reinforcement learning to trade China’s A-share market. They’re now running 2–3 million RMB in live trades. A conversation about data moats, the 500M RMB fund hurdle, and why the next DeepSeek might come from a quant trading desk.

By Dom Dotzauer — Hainan Airlines, 35,000ft — March 2026

Recorded mid-flight on Hainan Airlines — between chicken rice and coffee

Post-landing at the airport — L-R: JC (Cai Jincheng), Boris Ding (丁立, CTO, ex-Sequoia China), Dom, Richard

About AgentBull

AgentBull (agentbull.cn) — officially 煜马（深圳）数据信息有限公司 / Yu Ma Data — is a Shenzhen-based AI fintech startup founded in early 2025. Their tagline: “在不确定性中构建必然” — “In uncertainty, build inevitability.” Rather than fitting historical data, they use OSINT signals (hiring data, supply chain shifts, news sentiment) and multi-agent causal reasoning to produce forward-looking investment insights for China’s A-share market. The platform uses Qwen (Alibaba) as base models plus custom-trained sub-billion-parameter models with non-linear attention architectures (RWKV) for long-context financial processing. Launched at the Hengqin World Bay Area Forum in September 2025.

The Team

The company’s CTO and public face is Boris Ding (丁立), who led the data team at Sequoia Capital China (红杉资本), where he built knowledge graph infrastructure for VC deal sourcing. He’s the architect behind AgentBull’s multi-agent system and RLVR training pipeline. Wang Yuqiao (王玉乔) is the major shareholder and business lead, from a traditional family business background — he doesn’t work on the agent/AI side directly but manages partnerships and capital. Richard handles BD and international relations, and JC (Cai Jincheng / 蔡锦城) rounds out the core team. Domain registered January 2025, first Hugging Face model (CJK-Tokenizer) uploaded April 2025, product launched September 2025.

Why This Matters — Context

AgentBull sits at the intersection of two massive trends: China’s quantitative trading boom and the AI agent revolution. High-Flyer (幻方), the Hangzhou quant fund that built a 100 billion RMB (~$14B) portfolio using AI, spun off its research arm in 2023 — which became DeepSeek, the model that shook the global AI industry in January 2025. China’s quant hedge funds are now aggressively recruiting AI talent from US universities, and global firms like BlackRock are building ML models for the Chinese market. AgentBull’s bet: that a small, AI-native team with an ex-Sequoia data lead and OSINT-driven causal reasoning can carve out a niche — if they can solve the data access problem.

How It Started

Dom: Tell me about the company. How did AgentBull get started?

Richard: The founder is Wang Yuqiao — he’s our major shareholder. He comes from a family business background — traditional companies, not tech. He’s very young and got very interested in AI and trading. He wanted to combine both. He brought in Boris — that’s Ding Li, our CTO. Boris built the system architecture from scratch.

Dom: Which models are you building on?

Richard: We use Qwen from Alibaba as the base. Boris built our own smaller models on top — around 0.5 billion parameters. It took about three to four months for our first version, another three to four months of training to improve, and about two months ago we started running real money.

Boris adds (technical correction at the airport): “We primarily use Qwen. We also trained our own sub-billion-parameter models using RWKV-style non-linear attention architectures, mainly for handling long-context financial data. We use a multi-agent hybrid framework that chains different models together, where each model has different RLVR training objectives and reward directions.”

Editor’s note: AgentBull’s architecture is fundamentally different from how Western quant funds like Two Sigma use LLMs. Western firms typically use models as one node in an existing workflow — e.g., for sentiment analysis. AgentBull runs the entire workflow end-to-end with language models, with RLVR reward signals training across the full agent chain. Boris argues this means the alpha they can discover is structurally different: rather than traditional quant signals, they’re capturing “narrative economics” — identifying which market stories and narratives are dominant at any given moment, and predicting buy/sell decisions from the people who believe those narratives.

Real Money, Real Market

Dom: How much are you trading with right now?

Richard: We have 2 to 3 million RMB running in the market right now.

Dom: And the rest of the capital — you’re holding it back?

Richard: Yes. We’re not sure the strategy is good enough for the full amount yet. You try for two weeks, and if the strategy isn’t performing, you scale back, do more training, adjust the model. Then you try again. It’s iterative.

“You try for two weeks — if the strategy isn’t good enough, you scale down, train more, and try again.” — Richard, AgentBull

The Data Problem

Dom: What’s your biggest challenge right now?

Richard: Data. That’s the most important thing. We need high-quality trading data. If we had enough good data, we could put all our money into trading and make much more. But we’re a new startup — the big funds that have been around for years, they share data with each other. We have to start from scratch.

Dom: Could you build free tools for retail traders and get their data in return?

Richard: We haven’t tried this. But we need the data to be high-quality — trading decisions based on real rules, real strategies.

Dom: So you’re basically trying to deduce what trading strategies bigger traders are using — spy on their behavior?

Richard: We’d like to, but it’s very hard.

Boris adds (post-interview): “In many cases, data and compute are equivalent. With enough compute, you can obtain enough data — because the real bottleneck isn’t the data source. There’s massive amounts of OSINT and public data available, but it’s buried in noise. Separating signal from noise is fundamentally a compute problem. So our strategy for expanding our data isn’t to expand our data channels — it’s to get more compute, and to improve our compute efficiency: how to process more data per unit of compute and extract different signals from it.”

Fundraising & The 500 Million Hurdle

Dom: Who’s investing in you right now?

Richard: Family office investors. Wang Yuqiao has connections — they’ve known each other a long time. Most of them have family businesses, property, inherited wealth they’ve grown. For them, 5 to 10 million RMB is simple money. They give us a try, and if they’re happy, they give more.

Dom: Is there a turning point — a fund size where everything changes?

Richard: Maybe 500 million. That’s the hurdle. Once you’re at that size, other companies will take you seriously — potential data partnerships, cooperation. Below that, you’re too small for them to care. At 5 billion, you get easy access to everything.

Dom: Chicken and egg problem.

Richard: Exactly.

Editor’s note: This chicken-and-egg problem has a famous precedent. High-Flyer (幻方量化), one of China’s largest quant funds, built a 100 billion RMB (~$14B) portfolio using AI — then in 2023 pivoted its research team into a standalone AI lab: DeepSeek. The lesson: China’s quant trading ecosystem is one of the world’s most competitive AI training grounds, and the talent/infrastructure built inside these funds can spin off into globally significant AI companies. AgentBull is at the early stage of this same trajectory — the question is whether a small OSINT-driven fund can cross the data moat that separates startups from the Mingshis and High-Flyers of the world.

What Investors Want to Know

Dom: When potential investors hear about AgentBull for the first time — what do they ask?

Richard: First: the trading track record. Real results, not backtests. Then: what data do you train on, and how do the models work? After that, our background — what did the team do before? And then comes the casual talk. Getting to know each other personally. It’s very Chinese — you eat together, have drinks, build the relationship first.

AI Will Replace Work — But Not Yet

Dom: What should people know about what you’re doing?

Richard: AI is very powerful and will take over a lot of work. But honestly — not yet. AI handles code and office applications pretty well, but for real-world decisions with real money? It’s still not good enough on its own. We’re one of the first companies using AI models to trade in a real market with real capital. I think it’s a big opportunity for our generation.

Dom: Especially for people who are confident enough to learn AI and work with it — as a coworker, not a replacement. And that’s what you do, right? You combine human judgment with AI systems.

Richard: Yes, exactly.

“It’s a big opportunity for our generation — especially for those confident enough to learn AI and work with it.” — Richard & Dom, on the flight

Boris Adds: From Sequoia’s Radar to Trading Alpha

Dom: Boris, what did you actually do at Sequoia China, and how does it connect to what you’re building now?

Boris: At Sequoia we focused on the primary market — VC investment. We built a complex knowledge graph connecting startups across multiple data dimensions, structured as a network with companies as nodes. Through multi-hop signals — looking across different nodes — we could discover interesting patterns. For example: if we observed that nearly all SaaS startups were suddenly hiring implementation consultants at scale — not just one or two companies, but across the entire industry — that was a signal of an industry-level shift. We used something like a radar system, scanning this network to detect both macro industry changes and micro company-level signals. Like noticing a company showing sudden business growth based on certain indicators, meaning there might be investment value, and we could move first.

Dom: And the plan beyond A-shares?

Boris: Right now, processing market data is very compute-intensive, so we’re focused on A-shares. But from an algorithmic framework perspective, if compute and data sources can keep up, we’ll expand to futures, US stocks, Hong Kong stocks — all markets. We already have US and HK data, but we haven’t prioritized it. A-shares remain the core optimization target for our reinforcement learning pipeline.

“Data and compute are equivalent. The bottleneck isn’t the data source — it’s the compute to separate signal from noise in massive OSINT.” — Boris Ding (丁立), CTO, AgentBull

Arthur Wang Yuqiao (王玉乔), AgentBull’s founder and major shareholder, met the team at the airport. He wasn’t on the flight — he manages the business and investor relationships from the ground.

Epilogue

I was so deep in conversation that I walked straight past baggage claim and out of the airport with the team — all of whom had carry-on only. I had to run back through security to collect my checked bag. The security guard found this very funny.

Open Questions

Topics to explore in follow-up or Part 2:

1. What has the actual P&L looked like over the first 2 months of live trading? Win rate, drawdown, Sharpe?
2. What’s the regulatory landscape for AI-driven trading in China? Is the CSRC paying attention?
3. The data cartel idea — could complementary funds pool data for mutual benefit? Who would organize this?
4. What’s the path from 3M RMB to 500M? Realistic timeline and milestones?
5. Would international investors (EU/US) be interested, or is this purely a China-domestic play?

How This Was Made

This interview was recorded mid-flight on Hainan Airlines (Shenzhen to Shanghai route, March 2026) using voice memo, then transcribed and structured with Claude AI. Boris Ding added technical corrections and context in Chinese at the airport after landing — translated and integrated in real-time. The article was drafted, fact-checked against agentbull.cn, and designed during the time it took Dom to run back for his checked luggage. Final review happens via WeChat group with the AgentBull team. Human + AI workflow, start to finish — which is kind of the point of everything First Foreigner covers.