Original · Unique Research · 2026-07-10
Editor's note: This is the original Chinese author's interview with Fang Xiaolei and its framing. This English rendition retains the full text in source order, including all 15 Q&A items. The interviewee's product claims, scale figures, and market views are his self-reports, attributed to him and not independently verified. Person, company, and product names are preserved as source attributions.
AI Industry Observation
In the AI era, how exactly do we assess, develop, and supply talent?
An HR-background CEO on a tech-dominated AI track — on what grounds?
Two years on, Fang Xiaolei's AI Dexian Recruiter has passed 10 million cumulative interviews, iterating HR-vertical model capabilities on the self-built "Jinyu Superbrain" (近屿超脑) model backbone; clients expanded from 11 industries to 20-plus; the AI training business has served over 7,000 students. The company moved from "feeding itself" to HICOOL Global Entrepreneurship Competition second prize and "Go Global AI 100."
An entrepreneur with an HR background who spent 11 years in HR at big platforms and multinationals — on a tech-dominated AI track, on what grounds?
I talked with Fang Xiaolei for two hours, from business review to self-built models, from "black box" skepticism to organizational evolution. The biggest takeaway: Jinyu Intelligence's moat isn't model parameters, but an HR-background CEO's obsession with the act of "reading people."
Two Wheels: Training Isn't "Stealing the Show" but "Reading + Developing People"
In July 2024, Fang Xiaolei walked on two legs — AI interviewing plus AIGC training. Now both grow, but the structure changed. Someone asked him: training's share has grown; are you still an "AI interview company"? Will clients question your product focus?
"The recruiting side solves 'reading people'; the training side solves 'developing people.' AI interviewing lets enterprises know what kind of person they need; AI training lets such a person be cultivated."
Among the 7,000-plus students are university students, working career-changers, enterprise employees, and in-house trainees — four completely different groups pointing to the same question: in the AI era, how do we assess, develop, and supply talent? Plainly, this isn't "an interview company doing training on the side," but a move from a point tool toward an "assess-develop-match-deploy" data loop. The cash flow and frontline data training brings feeds back into AI-interview product capability.
From 6 Million to 10 Million: What the 7th-Generation Agent Changed
At the April 2025 Unique Awards, Fang Xiaolei said "nearly 6 million people have completed our AI interview." Now it's at the ten-million scale. The growth isn't a single breakthrough but several factors stacked: old clients expanding from campus recruiting to social recruiting and more business lines; new clients steadily added; usage depth per client rising. More importantly, the PLG model — letting enterprises and HR first try with a low threshold, then move from trial to formal purchase.
Industry coverage also widens. Beyond manufacturing, finance, internet, insurance, and banking, autos, accounting firms, supply chain, and new energy are expanding. High-end manufacturing shows clearer growth — such enterprises have stable hiring demand, clear job systems, and high demands on screening efficiency and interview consistency, where AI-interview value is easiest to see.
But the product upgrade most worth watching is the 7th-generation AI Dexian Recruiter Agent, on the self-built "Jinyu Superbrain" model. From L5 to L6, L7, the core change isn't a few more features, but the product moving from "can complete an interview flow" toward "more like a professional interviewer."
Two new capabilities in L7 technical interviews deserve special attention. The first is Code Review. The L7 AI Dexian Recruiter Agent combines the candidate's resume and project experience, auto-generates targeted code-review questions, and provides a complete code snippet generated by Jinyu's self-built model, asking the candidate to read, locate problems, and design optimization within a time limit.
This step doesn't only test whether the candidate finds bugs, performance bottlenecks, and security risks; more importantly it judges engineering literacy and project-governance ability, including whether they understand system design, can identify structural problems, and can propose actionable optimization. Compared with testing only code writing, Code Review is closer to real engineering and better identifies the engineering talent enterprises truly need.
The second is Vibe Coding AI collaborative programming. It focuses on whether the candidate can collaborate with a large model through prompts to complete real development tasks — requirement understanding, code generation, debugging and fixing, continuous optimization — judging their human-machine collaboration in the AI era.
Client feedback has also shifted clearly: in the past clients cared more about "can AI complete an interview"; now they care whether "AI can take on more complex, professional, real-business interview tasks."
Self-Built "Jinyu Superbrain": Not Because "We Also Have a Model"
Here's a counterintuitive choice. In 2024-2025, GPT-4, Claude 3, and Gemini leaped ahead. Everyone felt "isn't just calling the API great?" But Jinyu Intelligence stuck to a self-built proprietary model all the way.
"HR recruiting is special; you can't solve everything with a general API. An AI interview isn't just having the model ask a few questions and write a summary."
Behind it is how to understand job competency, judge candidate answers, dynamically follow up, score, explain results, keep evaluation standards consistent across candidates — plus the security of sensitive enterprise resumes, interview records, and assessment data. General-model risks are real: hallucination hurts scoring accuracy; insufficient understanding of some industry contexts; follow-up logic not matching real recruiting. Many enterprise clients have high demands on recruiting-data security and want a system they control.
So the self-built cost-benefit still holds. The core isn't competing with model vendors on general capability, but going deep on HR scenarios: more stable evaluation logic, more explainable scores, more controllable data security. Fang Xiaolei listed five things forming the moat: long-accumulated HR-scenario data, job competency models and talent-assessment methods, the AI-interview/follow-up/assessment/report product system, delivery experience across many real enterprise clients, and the assess-develop-match-deploy data loop. Even if big vendors open an HR-specialized model tomorrow, it raises the industry's underlying capability but won't automatically solve how enterprises use it, land it, and combine it with recruiting flows. Plainly, Jinyu isn't selling a model; it's selling the ability to "understand HR."
From "Black-Box Score" to "White-Box Logic": Evolution Forced by Clients
A 2023 CEIBS case recorded a real story: in 2021 serving Group A, a subsidiary's HR questioned the AI scoring logic. Black-box scores weren't trusted; system integration was hard. Fang Xiaolei admits they really faced many early doubts: why does AI score this way? Is it fair? Could a candidate be misjudged? Can business departments believe the result? These questions forced a key shift: from black-box scoring to white-box logic.
Today's 7th-generation product stresses "explainable scoring" — letting HR and business departments see where the candidate excels, where risks lie, what the score basis is, and what needs human follow-up. Client understanding is maturing. Few stay at the generic "is AI fair?"; more care whether scores are stable, evaluation logic is explainable, and results align with human interviews, re-interview performance, and hiring feedback. Jinyu's response is verifiable data: back-to-back human-machine comparison experiments, AI-interview results对照 with human interviewer assessments, then validating score stability and effectiveness against later re-interview performance. Fang Xiaolei always stresses a positioning: AI is a decision-support tool, not an absolute judge. For key roles, AI raises initial-screening efficiency and evaluation standardization; final hiring judgment still combines HR and business leads.
How Does a Stock Market Grow? Four Signals Not Depending on Macro Recovery
In 2024 Fang Xiaolei said "if the economy doesn't broadly recover, AI recruiting can't grow fast; we're still waiting for an opportunity." A year later his judgment changed: AI recruiting is of course affected by macro hiring demand, but it can't only depend on macro recovery. In a stock market, enterprises still have strong demand: cut costs, raise efficiency, improve judgment consistency, reduce ineffective interviews. More importantly, enterprises no longer need just an AI interview system but a whole solution around talent identification and organizational deployment.
Fang Xiaolei sees four signals, judging "the opportunity is coming": first, more enterprises aren't just trying AI interviewing but putting it into formal recruiting flows. Second, client needs shift from "interview a few candidates for me" to "optimize the whole recruiting flow for me." Third, both HR and business departments start caring whether AI-interview results are explainable, reviewable, and data-sedimentable. Fourth, clients expand from single-job pilots to more jobs and business lines. So growth isn't waiting for "hiring volume to grow" but deepening each client scenario. Jinyu has already derived service capabilities from system delivery: organizational architecture review, talent review, job competency modeling, interview-standard design. Not only delivering a tool, but being responsible for recruiting efficiency, talent quality, and organizational development outcomes. AI recruiting is moving from "optional tool" into "efficiency infrastructure."
An HR-Background CEO's Candor: Advantages Are Obvious, Blind Spots Hurt
Fang Xiaolei never shies from his non-technical background. The biggest help of HR background is understanding people, organizations, and jobs better. He knows which links in enterprise hiring hurt most, why HR is busy, why business departments are unsatisfied, why candidates have poor experience. This is the most precious starting point for making an AI-interview product. But blind spots are also obvious. Early on, he underestimated the complexity of technical productization, the difficulty of engineering systems, data systems, and product iteration. Running an AI company isn't only understanding HR but also technology, product, sales, delivery, capital, and organizational evolution.
After the company passed 100 people, he hit two typical pits. First, strategic decay: strategy is clear at the top but distorts as it reaches middle management and the front line. Second, cross-functional collaboration: tech, product, sales, delivery, training, operations each have their own language and goals; without clear mechanisms they easily work hard in different directions. I asked him: if you could choose, would you go back to 2018 "rejected by 69 VCs but product simple," or stay in 2026 "many clients but complex organization"? Without hesitation he chose 2026. "2018's product was simple, but uncertainty was huge. You didn't know if clients really needed it, or if the company could survive. 2026's problems are more complex, but they're the problems after growth. We have real clients, product validation, a team base, and a bigger market opportunity. Founders can't miss simplicity; they must learn to harness complexity."
Capital Is an Accelerator, Not the Engine
Near the end, I asked the question many founders care about: attitude to fundraising. Fang Xiaolei was direct: "open, but not anxious." After completing a ten-million-level Series A in early 2024, Jinyu has had no new funding news. But Fang Xiaolei says the company doesn't rely entirely on funding to operate. The AI-interview business has a stable enterprise-client base; the AI-training business brings more direct cash flow. "If there's suitable capital that helps us accelerate R&D, market expansion, overseas布局, and organizational building, we're of course welcome. But we won't fundraise just to fundraise. Startups ultimately return to customer value and business fundamentals."
"What truly decides a company's fate isn't one accidental opportunity, but long-term customer value, technical accumulation, and continuous delivery ability."
A Final Judgment
The AI-recruiting track is splitting. On one side are demo-type products — look cool but can't enter enterprise flows. On the other are products that truly root in enterprise workflows — explainable, reviewable, with data loops and repeat purchase. Fang Xiaolei's judgment: what disappears are products with only concepts, no real workflow, no repeat purchase. What truly remains are products that enter enterprise workflows, raise efficiency, explain results, and continuously review. Jinyu Intelligence chose the "dumbest" road: self-built model, white-box scoring, extending from interviewing to developing and supplying, deepening each client scenario. No chasing trends, no big stories. But in two years, ten-million-level interviews, 7,000-plus training students, 20-plus industries — these numbers say one thing. In the AI era, the company that best understands "people" knows best how to use technology.
More Conversation Details
Q1: Is Jinyu Intelligence still an "AI interview company"?
Fang Xiaolei: More accurately, Jinyu is moving from a single AI-interview product company to an AI-era talent-assessment, talent-development, and talent-supply platform. AI interviewing has always been our technical root and enterprise-client entry; it solves the core recruiting problem: how to read talent more efficiently, stably, and scientifically. But these two years the AI-training business has also grown fast. It's not a simple supplementary revenue, but the second essential demand we see in the AI era: enterprises and individuals both need talent who truly have AI application ability. So I won't define training as "icing on the cake" or say it "steals the show." More accurately, it and AI interviewing form our two-wheel drive: recruiting solves "reading people," training solves "developing people."
Q2: From nearly 6 million completing AI interviews to the ten-million scale now, where does growth mainly come from?
Fang Xiaolei: So far, cumulative AI-interview sessions have grown from nearly 6 million at the time to the ten-million scale. This growth isn't from one point but several factors stacked. First, old clients keep repurchasing and gradually expand AI interviewing from campus recruiting and some jobs to social recruiting, batch screening, and more business lines. Second, new clients keep increasing; more enterprises put AI interviewing into formal recruiting flows. Third, usage depth per client rises; AI interviewing is no longer just an initial-screening tool but combines with talent assessment, recruiting reports, ATS flows, and talent sourcing. Another important reason is advancing the PLG model. We open trials and free experiences to more enterprises and HR users, letting them start with a low threshold and truly see whether AI interviewing raises efficiency, improves experience, and lowers screening cost.
Q3: Which industries does AI interviewing penetrate fastest?
Fang Xiaolei: AI Dexian Recruiter now covers 20-plus industries. Beyond existing manufacturing, finance, internet, insurance, banking, it keeps expanding into autos, accounting firms, supply chain, and new energy. By penetration, high-end manufacturing grows most clearly. Such enterprises usually have relatively stable hiring demand, clear job systems and competency requirements, and high demands on screening efficiency, interview standardization, and hiring-decision consistency. So AI interviewing more easily enters high-frequency recruiting, higher-standardized job scenarios first. As PLG advances, it will also gradually expand from traditional large and mid enterprises' campus and social recruiting to more industries, jobs, and smaller enterprise clients.
Q4: Is the 7th-gen AI Dexian Recruiter Agent upgrade client-forced or proactive tech leap?
Fang Xiaolei: Both, but more accurately it's the joint result of real client scenarios pulling us and proactive tech upgrade. Clients force us to solve not lab problems but problems the recruiting front line meets daily. For example: can AI follow up? Understand a candidate's real experience? Recognize vague answers? Give assessment reports HR and business leads can read? Better integrate with the enterprise's existing recruiting system, online exam system, and ATS? Especially in technical recruiting, client requirements get more specific. Clients like Trip.com have many campus-recruiting jobs and high volume each year, with heavy technical-job screening pressure; relying only on traditional resume screening and human first interviews is hard to balance efficiency and precision. So in L7 technical interviews we strengthened two capabilities: one is Code Review, testing whether the candidate has engineering judgment and project-governance ability; the other is Vibe Coding AI collaborative programming, testing whether the candidate can collaborate with a large model through prompts to complete real development tasks. The 7th gen's biggest change isn't a few more features, but the product gradually moving from "can complete an interview flow" to "more like a professional interviewer" and "an AI recruiting Agent that understands real job scenarios better."
Q5: After the 7th-gen launch, what's the biggest change in client feedback?
Fang Xiaolei: The biggest change: people used to care more about "can AI complete an interview"; now they care whether "AI can take on more complex, professional interview tasks." After the 7th-gen launch, several changes are clear. First, interaction is more natural. Candidates don't feel they're just asked fixed questions by a machine; they feel the system follows up on their answers, and the interview is closer to a real structured interview. Second, reports are more usable. HR and business departments don't just see a score but the candidate's strengths, risk points, competency performance, judgment basis, and follow-up suggestions. Third, technical assessment is stronger. Especially in technical jobs, clients no longer settle for simple Q&A or basic coding questions but want AI to further identify the candidate's real engineering ability. Fourth, more interviewer feel. Some clients really say the 7th-gen experience is closer to a trained structured interviewer than a simple Q&A bot. Of course we won't say AI is fully equivalent to a human interviewer. More accurately, AI takes on more standardized, structured, high-frequency interview tasks, freeing human interviewers from repetitive screening.
Q6: Is the AI-recruiting opportunity no longer dependent on macro hiring recovery?
Fang Xiaolei: AI recruiting is of course affected by macro hiring demand, but it can't only depend on macro recovery. People used to think AI recruiting grows only when the hiring market recovers. But we now see that even in a stock market, enterprises have very strong demand: cut costs, raise efficiency, improve judgment consistency, reduce ineffective interviews, improve candidate experience. More importantly, enterprises need no longer just an AI-interview system but a whole solution around talent identification and organizational deployment. So the opportunity no longer comes only from "hiring volume growing" but from enterprises' whole restructuring demand for recruiting flows, talent assessment, and organizational deployment. AI recruiting is moving from "optional tool" into "efficiency infrastructure," and from point systems toward a "product + service + outcome delivery" integrated solution. That's the real opportunity we see.
Q7: Where is Jinyu's overseas expansion now?
Fang Xiaolei: Currently the business is still mainly domestic; overseas is in preparation and pilot promotion. HICOOL second prize and selection into "Go Global AI 100" will help our later overseas moves. On one hand it raised company recognition in the international innovation ecosystem; on the other it helps us connect overseas resources, partners, and local service networks. Our overseas move won't be a single model but "two lines in parallel." We won't simply copy the domestic product overseas; while keeping core technical capability and product advantages, we'll combine local legal compliance, cultural habits, recruiting flows, and candidate-experience requirements to redesign and deliver locally. Overall we believe AI-interview and AI-recruiting products have long-term opportunity overseas, but we'll advance in steady pilots, gradual replication, and gradual scaling.
Q8: Large models advance so fast; why does Jinyu still insist on self-built models?
Fang Xiaolei: For Jinyu, self-building isn't to tell a "we also have a large model" tech story, but because HR recruiting itself is special; you can't solve everything with one general API. An AI interview isn't simply having the model ask a few questions and write a summary. Behind it is how to understand job competency, judge candidate answers, dynamically follow up, score, explain results, keep evaluation standards consistent across candidates, plus security of sensitive resumes, interview records, and assessment data. So we don't set "self-built" against "external." We of course watch OpenAI, Claude, Qwen, DeepSeek, and use external model capability in suitable scenarios. But our core judgment is: what matters most in HR isn't whose general model has more parameters, but who better understands recruiting flows, job competency models, interview-evaluation logic, and real enterprise usage scenarios. Our self-build focus isn't competing with model vendors on general ability but going deep on HR scenarios: more stable evaluation logic, more explainable scores, more controllable data security, more stable and credible delivery to enterprises.
Q9: If big vendors open an HR-industry-specialized model tomorrow, where's Jinyu's moat?
Fang Xiaolei: Our moat isn't just the model but five things. First, long-accumulated HR-scenario data. Second, job competency models and talent-assessment methods. Third, the product system of AI interviewing, AI follow-up, psychometric assessment, and report generation. Fourth, delivery experience across many real enterprise-client scenarios. Fifth, the assess-develop-match-deploy data loop. If big vendors open an HR-specialized model tomorrow, it raises the whole industry's underlying capability but won't automatically solve how enterprises use it, land it, explain it, and combine it with recruiting flows. Jinyu's value is turning AI-model capability into a system in HR scenarios that truly lands, explains, delivers, and continuously optimizes.
Q10: How did the shift from "black-box scoring" to "white-box logic" come about?
Fang Xiaolei: It was jointly pushed by client education and tech evolution. Early on doing AI interviews, we really faced many doubts. Clients asked: why does AI score this way? Is this score fair? Could a candidate be misjudged? Can business departments believe this result? These questions are very valuable. They made us realize earlier that AI interviewing can't only pursue "automation" but must also pursue "explainable, verifiable, reviewable." So we gradually moved from black-box scoring to white-box logic. Today we stress letting HR and business departments see: where the candidate excels, where risks lie, what the score basis is, and what needs human follow-up. This isn't to please clients but the evolution AI must complete to truly enter enterprise key flows.
Q11: Is the AI Dexian talent-sourcing Agent extending from interviewing to headhunting?
Fang Xiaolei: The AI Dexian talent-sourcing Agent isn't a replacement or rename of the AI Dexian Recruiter, but a sub-product under its product system, an Agent capability we further extend around enterprise recruiting flows. Enterprise recruiting pain points aren't only in interviews. In real recruiting flows, resume screening, candidate communication, interview invitations, resume collection, status follow-up, ATS sync all take HR lots of time. The AI Dexian talent-sourcing Agent mainly solves these high-frequency, repetitive tasks in pre-interview and recruiting flows. It helps enterprises do initial candidate screening, automated communication, resume collection, flow follow-up, and system sync, freeing HR from lots of transactional work. So this isn't a simple brand repositioning but a feature extension and product deepening on the AI Dexian Recruiter's existing capability. It also isn't a simple replacement of traditional headhunting; more accurately it's an empowerment and cooperation relation. Traditional headhunters' value is industry understanding, client relations, high-end talent judgment, and complex communication; AI better handles high-frequency, repetitive, procedural parts.
Q12: Candidates use ChatGPT, DeepSeek to assist interviews; how do you prevent "AI interviewing AI"?
Fang Xiaolei: Candidates using AI to prepare interviews is already an irreversible trend. Our anti-cheating strategy doesn't rely on single detection but jointly solves it at several layers: question design, dynamic follow-up, multimodal recognition, and proctoring. First, on question design, we reduce standard-answer reuse through personalized questions, avoiding candidates memorizing templates or copying answers. Second, during the interview, AI deeply follows up on the candidate's answer, focusing on project details, behavioral details, decision process, and review ability. People who truly did the work usually tell specific details; answers temporarily generated or packaged easily expose logical inconsistency, missing details, or vagueness in continuous follow-up. Third, technically, we combine multimodal recognition, watching the candidate's answering state, expression rhythm, context consistency, and abnormal behavior, improving cheating detection. Fourth, in necessary scenarios we add physical anti-cheating means, like a second camera proctor. So future AI interviewing isn't simple "anti-cheating" but combining question mechanisms, follow-up mechanisms, technical recognition, and physical proctoring to better identify a candidate's real ability.
Q13: Is there real synergy between the AI-training and AI-interview businesses?
Fang Xiaolei: The AI-training business has now served 7,000-plus individual learners domestically. Learner profiles are diverse, mainly four types: university students, working career-changers, enterprise employees, and in-house trainees of enterprise clients. We don't teach training only concepts but around AI tools, AI Agents, AI application development, project practice, and enterprise-level scenario training, helping learners form AI ability that can be applied, practiced, and transferred to job scenarios. It has strong synergy with AI interviewing. AI interviewing lets us know what kind of person enterprises need; AI training lets us know how such a person is cultivated. Connecting the two is the loop from talent assessment to talent development to talent supply.
Q14: What's your current attitude to fundraising?
Fang Xiaolei: We won't disclose specific cash runway in public, but it's clear that Jinyu isn't a company that fully relies on funding to operate. Over the past years we've valued operating quality and cash flow. The AI-interview business already has a stable enterprise-client base; the AI-training business also brings more direct cash flow and market validation. This makes our fundraising mindset calmer. Fundraising matters, but it isn't the only premise of enterprise survival. My current attitude is: open, but not anxious. If suitable capital helps us accelerate R&D, market expansion, overseas layout, and organizational building, we're welcome. But we won't fundraise just to fundraise. Startups ultimately return to customer value and business fundamentals. Especially in today's environment, whether an enterprise can generate its own blood, serve real clients, and deliver continuously matters more than telling a big story. Capital is an accelerator, not the engine. The engine must be customer value.
Q15: If you gave enterprise HR, AI founders, and investors one piece of advice, how does AI truly enter the HR workflow?
Fang Xiaolei: Don't start from "who can AI replace," but from "which HR workflow is most worth AI-reconstructing." AI entering HR isn't first building a big system, but starting from concrete, high-frequency, repetitive, measurable scenarios. For example, campus-recruiting initial screening, sales-role first interviews, batch candidate communication, resume screening, talent sourcing, enterprise AI-capability training — these are good places to start. Enterprises shouldn't seek a one-step start; first pick a scenario with clear pain points, clear metrics, and business willingness to cooperate, and run it through. To judge whether a company suits starting with AI interviewing, look at three signals. First, is this role high-frequency recruiting with many candidates and heavy HR initial-screening pressure? Second, does this role have relatively clear competency standards? Third, is the enterprise willing to review AI-interview results against later recruiting results? If all three are yes, this enterprise suits starting with AI interviewing. Don't roll it out company-wide at once; first pick one job, one business line, one recruiting cycle, run the data, then gradually expand.