---
title: "Unique Interview | Let’s Talk Anti-Fraud Over New Year — Why a Baidu T9 Founded an AI Risk-Control Agent Startup in 2023"
author: "Unique Research"
sourcePublication: "Unique Research (FFCAP English library)"
originalPublishedAt: "2025-01-30T23:23:28+08:00"
canonical: "https://ffcap.cn/en/research/20250130a06hm600"
source: "https://ffcap.cn/en/research/20250130a06hm600"
language: "en"
cnSourceUrl: "https://view.inews.qq.com/a/20250130A06HM600"
---

# Unique Interview | Let’s Talk Anti-Fraud Over New Year — Why a Baidu T9 Founded an AI Risk-Control Agent Startup in 2023

_Unique Research · Interview · January 30, 2025_

## Founder profile

**Ouyezi** (handle) holds a bachelor’s from Tongji and a master’s from Peking University. After graduating in 2008 he joined China Construction Bank headquarters in application and security architecture. In 2016 he joined Baidu as Principal Architect (T9), chaired Baidu Cloud’s tech TC, and was the first product manager of Baidu PaddlePaddle AI Studio. In 2018 he joined Beike (KE Holdings) as senior director, building info-security and business risk-governance systems from scratch. In early 2023 he founded **Ouye Technology**; in August 2023 it raised an angel round from Apple Capital and is now in Pre-A, with Unique Capital advising.

## On founding

**Q: You were doing fine at big tech — why leave to start a company?**

**Ouyezi:** Looking back on 16 years — CCB, Baidu, Beike, Ouye — every move shared the same almost obsessive pursuit of “value.” First, personal value: how do you want to spend a life? Some coast, some grind, some speculate — preferences, not right or wrong. Mine is to struggle — from upbringing and because I want a life that means something. Second, value to society: does what I do add unique, incremental value, or just reheat what anyone can do? That pushes me toward new things.

From that lens I’ve been in a loop: enter a new environment as the weaker party, greedily absorb, grow until I can handle anything, then feel the value ceiling and move again. Maybe the end of that loop is founding.

**Q: Why AI risk-control agents as the direction?**

**Ouyezi:** AI is the biggest, best direction of the next twenty years; it will remake the world. Compare to PCs: we are in AI’s “1990s.” Early-’90s PCs were already celebrated high tech, but before mass adoption the internet swept the globe even faster. PC and internet pulled each other: booming apps drove PC adoption; better PCs fed internet boom. Which made which is a false question — they fused.

Today’s AI sits at the jump from “PC” (AI) to “internet” (AI apps). Over the next twenty years AI apps will rebuild industries one by one, with the same mutual pull: apps demand more from foundation models; model progress expands how far AI can remake industries. You don’t sit out a leap like that.

Why risk control? Past career — I believe in the 10,000-hour rule. To understand a field and talk with top people you need ~10,000 hours — five years at 5×8, or ~2.5 years at 6×12. Smart people can learn frameworks fast, but frameworks plus detail need time. I chose risk control because I know it; I don’t believe I can quickly master what others spent years not cracking. Dropping your strengths to compete with your weaknesses against others’ strengths is not rational.

Another reason: the domain could be huge, yet reality kept it small. Public “risk control” still means “click farms and promo abuse”; even some practitioners blur risk control and risk management. In fact every business has a shadow: loans imply loan fraud, insurance implies claims fraud, sales imply private deals, logistics imply rough handling, traffic lights imply red-light running. Not every scene needs intervention. I prefer “healthy enterprise operations”: anything that seriously harms the core business through organized human malice — spanning C-side risk, B-side internal control and quality, G-side compliance, even state-level confrontation. Anyone who talks risk control without the core business sinks into governance for its own sake — which is why the industry never scaled. Demand exists; supply missing suppressed it.

Long term, AI as foundation and “healthy enterprise operations” as entry is our path. Vertically in risk management we can remake a large market; horizontally we are a data-and-AI adversarial company — tech is general — so expansion options exist. Today’s job is survive so we can choose later.

**Q: Thoughts after a year-plus of founding?**

**Ouyezi:** Stress eating is real — Hong Kong dramas weren’t joking. Jokes aside, the worst part was weight; the best was ending mental churn over “value fit.” I’ve thought a lot about human nature, orgs, capital, business, enterprise services, institutions, professionalism. Feelings on the data mega-track are the richest and most lasting — a huge topic.

Data is today’s biggest factor of production; data flow and use are the biggest relations of production (without flow there is no LLM, no Sora). In banking, cognition and action clash hard: twenty years of digitization, >95% off-counter business, massive collection — yet internal flow is tightly controlled and data barely gets used. Causes span bank mindset, architecture, IT orgs, security, departmental power — splitting governance rights from usage rights. HQ wants governance, branches want use; tech wants governance, business wants use. Until that bottleneck breaks, effort is wasted.

On the vendor side most paths go bottom-up — selling ML/DL platforms, engines, models, methods. Tech can’t catch; business feels nothing; without tech–business linkage bank data and AI don’t land, so flow never forms from innovation. In banks, business development must lead; data flow and tech progress support. Most vendors lead with tech, assume data can flow, wait for business impact — which fails, because what’s blocked is not technology but the tech org’s power position. How can a vendor push the buyer’s org chart? So ML/DL not scaling in banks is unsurprising.

Better path: tools business teams can use directly — wrap models, algorithms, and data into out-of-the-box tools for operators, not only strategy researchers or data scientists. Solve business problems to drive data flow; flow drives tech that feeds business — a loop. I increasingly believe that, without changing org charts, is the only way bank big data becomes intelligent. Some call it scene know-how. On the surface we solve risk scenes; underneath we are building analytics products that fit SOE org structures.

## On risk control

**Q: People see risk control as a blocker — what is it really?**

**Ouyezi:** Hard-to-define concepts are ones everyone can chat about. Risk practitioners talk past each other like reading Japanese kanji as Chinese — familiar shapes, different meanings. Zhang’s “risk” and Li’s “control” differ in scene, tech, even philosophy. Efficient talk needs standard definitions — which is hard.

_Narrow_ risk control usually means control means — process and tech. Defaults differ: internet teams mean tech controls; internal-control teams mean process.

_Broad_ risk control is the full cycle: assessment/exposure/identification, confirmation (tied to risk appetite), narrow control, and tracking. Domain theories differ in wording, not essence. Identification is the premise of control — but not all risks are clear, and not all identified risks need control. That sounds empty; it is a common enterprise mistake.

**Q: How do you classify risk control — main scenes?**

**Ouyezi:** Risk follows business; business is too diverse for simple taxonomies. Mature industries have domain definitions. Banking, under Basel ideas, names credit, market, liquidity, operational, and other risks. Internet definitions are freer and messier: account security, click farms, promo abuse, content safety — or “anti-cheat / anti-fraud / anti-black-industry.”

I try a role frame: **C** customers; **b** employees (sales, ops, staff, managers); **B** the enterprise and its decision-makers; **G** regulators and government. Almost every risk event is interest invasion among these four. Examples: **C2B** — customers harm the firm (promo abuse, loan fraud); **C2C** — customers harm customers (telecom fraud); **b2b** — employees harm employees (poaching deals); **b2B** — employees harm the firm (private orders, embezzlement).

Other high-pain, under-solved scenes: commercial espionage (B2B) — a real-estate firm’s weekly management report of deals and visits appeared on a competitor’s desk the same morning, every week; organized, interest-driven. Quality management (B2b) — not invasion but preventing it — is acute in service industries: in some labor-intensive sectors ~3% of staff have criminal records (mostly drug/economic; some violence). Guaranteeing service quality means managing people; managing people starts with hiring background checks.

This classification’s first lens is who gains and who loses. Only the damaged party has enough drive to control risk. Some call that risk control’s first principle: attacks happen because attackers gain; control happens because victims lose. Tech and product matter, but only under the damaged party’s drive. Two scenes that often fail: telecom fraud (C2C) — platforms, carriers, banks are not the primary losers and may even gain indirectly; without the 2022 Anti-Fraud Law, voluntary drive was weak — victims are easy to blame, banks hard to clear. Second, marketing acquisition abuse: risk sees fake acquisition; marketing sees higher CAC if targets are hit — whether the firm loses is a CEO question. Designing against human nature fails; good systems fit it. Common root: unclear damaged party → weak drive. Who hurts, who fixes — taxonomy should find who truly hurts.

**Q: What are the main schools of risk-control tech?**

**Ouyezi:** Three: prediction, evidence, and defense-in-depth.

_Prediction_ — oldest and most mature in tech/product. Find features and anomalies. How do you spot a robber in a crowd? Expert rules say “stocky, knife, stocking mask” — and fail on edge cases. AI softens enumeration and even finds features experts can’t explain but that work. Still: features, profiles, tiers, scores — correlation more than causation, while ops wants causal explainability. In plain words: “You look like a bad guy” — and I won’t explain why.

_Evidence_ — old idea, blank product space. Trace to find proof. How do police crack theft? Clues (weapon sold at one shop), suspects (few buyers), fingerprints. Plain: “You _are_ the bad guy — only you link to this.”

_Defense_ — security’s defense-in-depth: auth and permissions woven into process so one breach doesn’t take the system. A bank vault needs more than one door. Raise the cost of evil; attackers fear not whether they can break a gate, but whether success brings surprise or disaster.

## On bank risk control

**Q: What is distinctive about bank risk control?**

**Ouyezi:** Banks prioritize transaction risk over marketing fraud. Order: credit risk > transaction risk > telecom-fraud risk > marketing risk (Basel market risks like FX are out of our scope here).

On scenes: for transaction risk the floor is strong auth — experience may suffer, but function stays. Models and algorithms are icing, not the cake. Why push second-gen USB tokens? Security — and so customers cannot deny they transferred funds.

On tech: bank stacks grew from 1990s financial IT — a different tech tree from internet big data, with different product roles, business–tech relations, and ops. That is why ML/DL struggle to scale on core bank business — worse at larger banks.

On ops: banks need strong explainability; anti–telecom-fraud is no exception. Internet acquisition lets the host own interpretation rights; customer funds’ interpretation rights belong to the customer — which hardens ML/DL use.

On data: not shortage but power fragmentation — HQ vs branches, business vs tech, even among tech teams — old entitlement maps, hardest at big banks. Against telecom fraud the fight can look like Ming elites vs Manchu cavalry: brilliant, well-resourced teams beaten by rough, profit-driven gangs. Long processes and crossed org lines vs fast feedback and interest-aligned crews — losing is unsurprising.

**Q: So “tech doesn’t get business, business doesn’t get tech, walls are high”?**

**Ouyezi:** Yes — and that architecture almost cannot change. Stating the truth sounds like complaining; only honest diagnosis yields the right medicine. Vendor opportunity lives in these absurd realities. Products that solve the abstract problem will change the industry — not via relationships or regulatory shortcuts, but by hard work on surface scenes and underlying orgs. Very hard — and a huge commercial chance. One reason I left to found.

**Q: As licensed credit and risk institutions, what is the history of bank fraud?**

**Ouyezi:** Fraud and anti-fraud are eternal in finance. Attacks iterate; defense stays unfinished. Online fund fraud against banks has had roughly four waves in twenty years.

_Pre-wave (before 2011):_ as banks went online, early e-banking fraud appeared — Trojan transfers, phishing. Big four banks handled ~300–500 phishing sites a year. Sporadic and probing; no nationwide playbook yet.

_First wave:_ Trojans against first-gen USB tokens (2011–2013). SMS blasts saying the token needed “upgrade”; victims plugged in and ran malware that used OS API gaps to call the token’s certificate libraries. Ebbed after second-gen tokens ~2013. Strongly technical attacks.

_Second wave:_ credential stuffing and card skimming (peak ~2013–2014). Stuffing rode mass leaks — e.g., a major developer community’s 6.5M accounts in Dec 2011 “opened Pandora’s box,” followed by leaks at many large internet firms (hundreds of millions of accounts across brands). Criminals stitched username–password pairs and automated logins at ~11.7% hit rates, then long-tailed victims. Skimming — ATM/POS capture of PIN and track-2, then white cards — is older offline craft. Crime shifted from pure tech toward scam-plus-tech, still tech-dependent.

_Third wave:_ fake base stations (~2015). Phishing sites jumped from thousands a year to thousands a month. Fake SMS from “bank CS” pushed fake apps (not always classic Trojans); victims leaked mobile-banking credentials; phone social engineering harvested SMS OTPs. Less reliance on deep tech.

_Fourth wave:_ telecom fraud explosion after ~2016 — marked by the September 2016 Xu Yuyu case. Ever since: always governing, never fully containing. Attackers largely abandoned tech for pure social engineering. The oldest forms often last longest.

**Q: Sounds like “as virtue rises one foot, vice rises ten”?**

**Ouyezi:** Yes. Confrontation never stops. Bank controls get better and harder; crime moves toward simple, crude, effective. Liability shifted from “bank at fault” toward “customer at fault,” so fund-security work carries more regulatory and social-responsibility weight. Across eras one premise stays: cross-domain customer data leaks. Knowing history may not directly stop the next attack, but fraud vs anti-fraud is people vs people — only knowing the adversary’s past helps judge the future.

## On bank anti-fraud

**Q: What drives banks to fight fraud?**

**Ouyezi:** Some say social responsibility; some say regulators need a handle. Deeper regulatory drive is capital outflow from telecom fraud — analyses put transfer scale on the order of **¥500 billion**. Not only livelihood — national economic security.

**Q: What does bank anti-fraud look like today?**

**Ouyezi:** It is not a cat-and-mouse game; it is wolves eating sheep. Wolves are money-laundering rings (not always the scammers); sheep are banks. Strategy: make your sheep thinner than the next bank’s — so thin the wolf won’t bother. In plain words: mass limits, freezes, account friction — racing who can make experience worse so criminals don’t care to use you.

But there is always a thinner sheep. Wolves stop when hunt cost exceeds reward, not when every sheep is skinny. When all are thin, wolves still find the fat among the thin. Fat vs thin is relative; eating is absolute.

Positively: 2023 was a peak year; under regulators, banks fought at all costs as a political task — necessary and effective to blunt the surge. Without endogenous drive, durable mechanisms, and system planning, the wolf-sheep game continues until banks build anti-fraud operating frameworks (org and process). Anti-fraud is not in a steady state — it is just starting. Banks rarely leave projects unfinished, and professionally there is much still to do — so this market is larger than most think. I’m firm on that.

**Q: What traps banks in anti-fraud?**

**Ouyezi:** Objectively it is risk management blending business ops, black-industry confrontation, data mining, models, and risk ops. Classic risk-management mistakes show up widely: treating it as temporary; treating it as one department’s job; “just ship feature X” (non-pro thinking); or doing everything at once (no quantified risk assessment).

**Q: What’s the way through?**

**Ouyezi:** Fraud vs anti-fraud is sustained human confrontation. Only an armed anti-fraud ops team can face organized crime. Two layers: people vs people — no fancy model outruns someone who fails or gets killed. And “confrontation” is not flattening criminals; it is making them find your sheep too annoying to hunt — so they go next door.

So the only path around that essence is a full-time professional anti-fraud ops team — preferably but not necessarily a hard org, always with absolute power to marshal resources. Nationally the anti-fraud center already runs as an entity; in 2024 some large commercial banks began entity-izing teams. Whoever builds first leads the race. Entity orgs are a market milestone: only dedicated people build professional process and systems — and standard products and services are required elements of those systems.

## How Ouye does bank anti-fraud

**Q: How does Ouye plug into bank anti-fraud?**

**Ouyezi:** Our products center on _operations_ — perhaps the only vendor that leads with ops. Ops here means two cores around reconstructing attacker processes: black-industry confrontation and case reconstruction.

_Black-industry confrontation_ answers “unknown attack, how defend?” Crude analogy: to not get hit, know how they hit. Many intel vendors exist; in bank anti-fraud the gap is turning intel into usable bank controls. Hard — but productizable; in about a year nearly a hundred banks and police units used ours, many inbound trials. Mining depth is still limited: intel people don’t know bank business; bank people don’t know intel workflows.

Important: even perfect black-market intel covers only ~10% of involved cards. Greater value is completion and high-level breakthroughs — special-forces raids, not mass charges. Betting all on intel is sending special forces into frontal assault.

_Case reconstruction_ uses the bank’s own data to analyze involved-card behavior, find pre-crime signatures, and interpret them as “crime business scenes.” Still know-attack-to-defend; difference is noise of normal behavior — needle-in-ocean for the inexperienced. Some get it in a week and are experts in six months; others never crack it after years. Overall it is a trainable craft, especially with usable tools.

A reconstruction outputs who, where, via which channel, into which system, did what distinctive act. Aggregate: which attack scenes hit this bank, and at what share. That is model-building input. Banks today don’t lack model R&D, decisioning, or compute — they lack model-building thinking. Like building great assembly lines with no feedstock. Reconstruction is feedstock — our entry. Multimodal large models help a lot on collection, processing, and business analysis — things we wanted but couldn’t do before.

Short version: banks may not build their own ops teams, but they cannot escape needing ops. Vendors win by helping buyers build runnable, sustainable confrontation mechanisms. Our team are not gray-haired theorists; we are veterans from risk scenes. We don’t do pretraining — we build agents. I’m not alone: Meituan’s L10, a Baidu T7 who rose four levels in five years, 15+ year bank-data veterans. Timing feels right.

**Q: After a small entry, how do you win more business?**

**Ouyezi:** In the new economy, compliant operations and cost efficiency are hard needs. Enterprise risk scenes are strongly correlated. Our vision: on the healthy-enterprise-operations mega-track, on a financial-risk AI-agent product base, solve operating risk faster and better — anti–bad debt, anti–illegal lookup, AML, anti–data leak, anti–fund diversion, and more. Beyond banks and police, large enterprises and sectors (military, national security, and others) need help.

— END —

---

Editorial English edition published on UniqueCapital: https://ffcap.cn/en/research/20250130a06hm600
On-site reading page: https://ffcap.cn/en/research/20250130a06hm600