Image generated by AI
TMTPOST -- On the screen flashed "367."
It is a quantity artificial intelligence (AI) recommended for restocking a specific shoe model in a particular region over the next 30 days. This number, generated by Belle Fashion Group's AI-powered platform Li Ying, represented the culmination of 200+ data dimensions: market trends, procurement cycles, city-specific sizing patterns, even local weather.
Standing beside me was the head of one of Belle Fashion's brands. I leaned over and asked, "Will you follow the AI's restocking suggestion?" His response was pragmatic: in the short term, they'd reference it, but long-term, they'd fully rely on AI.
Li Ying is a data- and AI-driven product operations platform co-developed by Belle Fashion and DeepTech. Ji Yanli, Vice President of Belle Fashion Group, explained: "Li Ying's core strength lies in process visualization. Previously, operations were a black box—we lacked a complete way to track product flows. Li Ying reveals the entire product lifecycle, not just inventory. It shatters traditional business intelligence (BI) boundaries by dynamically integrating time, space, stock, sales, and restocking into a single interface, creating an interactive 'data analysis film' that boosts efficiency. It harnesses AI to make data dynamically generate business value."
From Concept to Reality: The Birth of Li Ying
The project was launched in 2021. During my first demo viewing, I watched terabytes of item-level data stream across a massive screen, meticulously organized chronologically. The phrase "white elephant" came to mind—something grand yet seemingly impractical at first glance.
Ji is a retail veteran. Despite his background in business, he is obsessed with data, management, and organizational efficiency. He believes large-scale offline retail is inherently complex, requiring a unified communication mechanism to ensure precision across sprawling operations. That universal language for it is data—and the rules governing data.
He often recounts his days as a regional head: initially, solving a city-level issue meant flying there, wasting a few days to align on solutions. Later, he implemented a layered Excel system, standardizing data logic. Eventually, discussions were reduced to a simple call: "What's the issue in row X, column Y of sheet XX? Your plan? Okay, goodbye." Consensus became effortless.
Belle Fashion's digital foundation was strong—it built its own ERP system as early as 2014. But as online and offline operations merged, legacy systems faltered. In 2020, it began constructing a middle platform, restructuring around four pillars: organization, authority, business, and communication—forging a collaborative network.
Ji's philosophy is blunt."Digitalization is a management model—data-driven resource allocation. For retail, it means real-time clarity: How much stock? How many people? Where's everything? That's digitalization, plain and simple," he said.
By deconstructing siloed systems and reorganizing functions around business scenarios, they aimed for seamless collaboration—where systems find people and data finds people. This required a centralized data repository: a big data platform. Belle Fashion adopted a Lakehouse architecture to unify data services, standardizing data, metrics, and applications—a monumental technical and managerial feat.
It was amid this effort that they crossed paths with Zhao Jiehui, the CEO of DeepTech, in 2020.
DeepTech's Pivot From Middle Platforms to AI-Driven Solutions
Zhao is an industry veteran. After 11 years at Huawei leading core router teams, he joined Alibaba Cloud during the internet boom, overseeing enterprise services. Restless for change, he believed enterprise transformation through internet technologies was inevitable. He set up a team and dove into entrepreneurship.
His timing was fortuitous. The "data middle platform" concept was peaking, and DeepTech—armed with Alibaba Cloud expertise and strong execution—quickly rose to prominence. At a TMTPost summit in Sanya, their middle platform forum was packed beyond capacity. Success brought clients and investors, including Hillhouse Capital, Monad Ventures, and IDG.
But reality soon struck. Zhao realized middle platforms struggled to deliver tangible business value. "Internet giants need shared systems for massive scale, but most B2B firms don't. Data middle platforms offered tool-like utility but lacked depth. When I saw their limited impact, we pivoted fast."
Shareholders praised Zhao's willingness to admit mistakes and adapt. By 2020, DeepTech halted new middle platform contracts, focusing instead on flawless delivery of existing projects. "We weren't visionary—just desperate. A startup's margin for error is zero."
Two paths emerged: move upward—integrate data analysis with operations to drive value; or move downward—build a next-gen Lakehouse foundation.
Inspired by Databricks' Lakehouse architecture (merging data lakes and warehouses), DeepTech benchmarked against it, developing a full-stack data platform. Their agility stemmed from a unique R&D ethos: the CTO's primary role wasn't daily development but tracking emerging tech. This allowed early exploration of NLP, large models, and Lakehouse systems.
Yet, data platforms alone weren't enough. "Even with perfect data processing, it's just left-brain analytics. Without understanding business intent, numbers are lifeless."
Then, the stars aligned—Belle Fashion entered the picture.
Belle Fashion's Data Challenge: Six Warehouses, One Solution
Snowflake CEO Sridhar Ramaswamy once said, "The real value lies in companies that own customer relationships, deliver clear value, and embrace AI to avoid disruption." Belle Fashion fit this perfectly—but its data was a mess.
Since 2014, Belle Fashion had built an ERP system covering merchandise, procurement, inventory, sales, and finance. But rapid growth spawned siloed data warehouses—six in total—with redundant processing and inconsistent metrics. Changes required updates across all warehouses, creating chaos. A two-year data governance project consolidated them into one unified platform.
Belle Fashion went through a three-phase data platform journey: launching a project codenamed “Business Data Dictionary” to standardize chaotic data and improve usability;Partnering with DeepTech to build FastData, a real-time Lakehouse; creating Li Ying, a lifecycle visualization tool for products.
Today, Belle Fashion's big data platform processes 3PB+ across hundreds of servers, with all governance on a single platform—boosting efficiency and scenario adaptability.
The Partnership: Data Meets Scenario
DeepTech's entry was serendipitous. Zhao recalled, "We'd just launched our lakehouse product. Belle Fashion had no prior exposure but understood its utility." Traditional IT systems created barriers; Belle needed microservices aligned with business scenarios, demanding a new data architecture.
Through discussions, Zhao realized Belle offered something rare: real-world scenarios. Ji later recounted their negotiation: "Lao Zhao said DeepTech had funding but needed challenging, high-value scenarios. I had both."
Zhao approached the collaboration with zeal: "I wanted to create something remarkable. Without tackling complex scenarios, you're just another vendor."
Li Ying broke technical barriers. Traditional BI dashboards handled kilobytes; Li Ying processed terabytes, interfacing directly with big data systems to render TB-scale data in seconds—a 10,000x leap.
In 2021, DeepTech joined the Li Ying project proactively. Early iterations were grueling: initial full-data runs took 42 days, later reduced to 5. But the effort paid off—Li Ying's incremental updates now run every 0.5 seconds.
Yet, Zhao saw limitations: "Even with perfect data, without understanding business intent, it's just a fancy report."
The AI Advancement: Left Brain + Right Brain Logic
By 2022, Belle and DeepTech faced a new challenge: Li Ying had hit its technical ceiling. The broader market struggled too—data platforms were built, but their value remained unclear.
At the time, NLP and ABI were gaining traction and seemed practical for applications like conversational chatBI. We assembled a team to investigate their potential in reducing service costs and improving efficiency. However, we soon faced challenges—deriving real value from enterprise data required complex business logic. Simple data acquisition and statistics were insufficient; true value emerged only when integrated with business task identification and decision-making processes. Without this, the system was merely a reporting tool, sometimes even less advanced than existing solutions.
During this period, NLP proved inadequate for recognizing business needs beyond simple data analysis. The core issue was the ability to flexibly identify deep data platform use cases, business task logic, and align them with enterprise operations. At that time, it was unachievable. However, the emergence of large models changed everything. With the release of GPT in late November, we initiated research, and by our strategic seminar in March, we committed to forming a dedicated team. Initially centered around GPT, we later shifted focus to Llama2 and embraced the broader open-source model ecosystem. This breakthrough clarified our challenges and provided solutions.
We realized that beyond data analysis, deep business integration was essential. The data platform required substantial upgrades—not only for structured data but also document-based knowledge and logical data. These enhancements transformed our product system, demanding seamless integration between product development and sales.
The previous data platform had reached its limit in processing business data. While structured data governance was well-established, the evolution of AI demanded greater knowledge integration, encompassing business logic, images, and additional information streams. The challenge extended from data storage to analysis and governance, requiring a unified multi-modal metadata system, automated intelligent recognition, and information extraction. Data synthesis was also necessary to generate high-quality integrated data for large models, necessitating a significant technology stack upgrade.
These insights led to a major enhancement of DeepTech's data platform, culminating in the FastData Foil version. This upgrade enabled unified management of multi-modal data while maintaining compatibility with traditional data applications and tokenized data for model training.
The "Heavy Armored Brigade" Strategy
DeepTech's corporate culture incorporates military-inspired terms, influenced by Zhao's passion for military strategy and Mao's writings. One such concept, the "Heavy Armored Brigade," was introduced during DeepTech's uncertain period in 2019. It established an organic cycle of technology pre-research, development, and sales.
"Heavy Armored" refers to leadership-level engagement. Zhao mandates that all R&D team leaders and executives must visit potential and existing clients by the first quarter after Chinese New Year. These visits are not focused on project sales but on strategic benchmarking—aligning industry trends, products, and business logic with clients. This initiative has yielded two key benefits:
1. Assessing whether frontline sales teams are effectively engaging with clients.
2. Enhancing DeepTech's competitive advantage against major corporations by leveraging senior leadership expertise.
By fostering close client relationships and iterative feedback, this mechanism has improved customer service quality and strategic alignment. R&D investments in this initiative allow DeepTech to integrate client insights into their development process seamlessly.
Strategic Benchmarking and AI Implementation
At the 2025 Strategic Benchmarking Cooperation Kick-off Meeting in Shenzhen, Belle Fashion and DeepTech showcased their collaboration. They introduced the Deepexi-RM reasoning model and FastAGI, demonstrating AI applications across Belle Fashion's commercial operations. These included the BaiMing AI Data Query Bot, BaiCe AI Analyst, Li Ying AI Advisor, Single Product Operation Brain, and BaiGuan AI Assistant. The event also outlined the 2025 framework for deploying industry models and Agentic AI applications.
The AI applications highlighted the rapid progress of both companies over the past two years, establishing effective methodologies in model engineering and AI implementation. The collaboration proposed a structured "business-system-data-model” capability loop:
- Business operations determine system functions.
- Systems generate structured data for business logic.
- Data warehouses integrate business processes, forming datasets.
- Models leverage these datasets to optimize decision-making.
- AI-driven decision-making flows back into business operations for execution.
To move beyond basic data analysis and improve AI-driven business strategies, DeepTech and Belle Fashion developed an innovative "Left-Brain and Right-Brain” logic. This architecture became central to Deepexi's enterprise large model framework:
- The left brain (algorithm) focuses on structured data (e.g., sales forecasts, inventory statistics).
- The right brain (semantics) interprets unstructured data (e.g., consumer sentiment, competitor analysis).
Using SFT+RL training, the large model learns the underlying business logic behind data patterns, ensuring strategies align with both quantitative insights and business realities. This approach enables a shift from local optimization to global intelligence and from passive response to proactive decision-making.
For example, if the left brain detects a 20% decline in inventory turnover, the right brain contextualizes external factors such as weather conditions or competitor promotions. The AI system can then autonomously recommend solutions, such as adjusting product allocations or launching counter-promotions. Previously, this process took 1-2 days due to manual coordination; now, AI enables real-time strategy execution.
Deepening Collaboration and Scaling AI Solutions
Deepexi and Belle Fashion have collaborated for five years, making Deepexi the company's longest-standing AI partner. Despite occasional strategic disagreements, their relationship remains strong, fostering joint problem-solving and continuous innovation. Employees even adapted the lyrics of James Wong's classic song A Laugh in the Sea into A Laugh in the Data to reflect their shared journey.
Ji, a key stakeholder at Belle Fashion, emphasized the balance between customization and scalability in their partnership: "Excessive customization isn't necessary. If a feature aligns with your product vision, implement it; if not, we can manage it ourselves. Our collaboration with Deepexi works because they evaluate our needs and create solutions that exceed expectations."
Ji also advised Zhao on refining the Li Ying solution, noting that its complexity hindered adoption. He suggested leveraging large model technologies to improve usability and showcase DeepTech's expertise. This feedback prompted Zhao to rethink product scalability, leading to the development of the Deepexi Enterprise Large Model and FastAGI.
With the upgraded FastData Foil and enterprise AI models, Deepexi has rapidly expanded its client base in the large model industry. Reflected on the transformation, Zhao said:
"Unlike other partnerships where we implement IT managers' directives, our work with Belle Fashion is about creating business value. Their ability to lead product innovation aligns with our agility. AI and next-generation data platforms allow us to integrate enterprise knowledge deeply, offering intelligent applications that redefine China's B2B market. Every technological shift brings both risk and opportunity. Thankfully, we've successfully navigated two production cycles, ensuring our survival and growth. China's B2B ecosystem needs this evolution."
更多精彩内容，关注钛媒体微信号 （ID：taimeiti），或者下载钛媒体 App
