Leading Party building and tackling challenges with intelligence: a breakthrough path for technological innovation in a grassroots customs branch

At around 9pm, the business hall of the Yantai Customs Port Office under Qingdao Customs was still lit up, and strings of codes were rapidly scrolling on the computer screen. Active Party member Luan Yanchang stared closely at the screen, repeatedly debugging the algorithm model he had just written. Beside him, the branch secretary Yang Zhenlei is carefully checking each row of data with the results calculated by the model.

They are tackling a long-standing problem that has plagued customs - intelligent prevention and control of tax risks for imported cosmetics.

Qingdao Customs, as the main import port for cosmetics from Japan and South Korea, imports over one million customs declaration documents annually. Under the traditional regulatory model, customs officers need to manually review prices, classifications, and consumption taxes item by item, which is inefficient and difficult to fully cover.

At a branch business analysis meeting, Luan Yanchang proposed the idea of "using technological revolution to change the tactics of the sea of people": "Using AI models to automatically identify the brand and model in the customs declaration form, and then comparing it with the backend price database, so as to accurately identify risks. ”

From "0" to "1": The First AI Risk Control Model to Cross the Technological Divide

In the early stages of project initiation, technical bottlenecks emerged.

Although the branch members have some basic knowledge of Python and SQL programming, embedding them into business systems requires the use of HTML and JavaScript languages, which is unfamiliar to most members of the team. The branch has enabled young people to take on the main role and set up a "Youth Party Member Challenge Project", forming a research and development team with Luan Yanchang as the technical core and the collaborative support of all party members.

During that time, we would review documents during the day and become 'programmers' gnawing on code at night, "he recalled. Yang Zhenlei, who has a certain programming foundation, generously shares his experience with him, and the two often discuss and write code until late at night. Other party members are responsible for the standardized processing of massive historical data and the refinement of business rules.

After three months of concentrated efforts, the first version of the intelligent risk prevention and control model for cosmetics was successfully launched in November 2024. This first generation model has been able to automatically intercept some obvious abnormal declarations through a rule engine, freeing customs officers from tedious preliminary screening. However, there is still significant room for improvement in the recognition accuracy and coverage of the model, and it is still powerless for complex declaration content.

The real technological breakthrough will occur in early 2025.

When large-scale AI models such as DeepSeek emerged, Luan Yanchang keenly realized that deep learning technology could be the "key" to completely solve this problem.

Self Evolution: When Customs Operations Encounter the Cutting edge AI Technology

Traditional rule-based programming teaches machines how to do things, while AI deep learning allows machines to learn how to make judgments on their own. ”Luan Yanchang reported new ideas to the branch.

This time, technological breakthroughs have entered the deep-water zone. From model deployment to prompt writing, from JSON data processing to result recursion, every step is an unknown domain. The branch has launched a crazy self-learning mode. They deployed a neural network model specifically designed for the fusion of customs declaration text and multidimensional data using the latest open-source AI framework. This model can understand the natural language of product descriptions, recognize "30ml/bottle+20g/vial+15g/tablet" and calculate the total net content.

After 8 months of iterative optimization, by September 2025, the recognition accuracy of the model has jumped to 90%, with an effective rate of up to 97%.

This means that out of the 1 million cosmetic declaration data each year, over 970000 can be accurately screened by AI, the coverage of risk identification has been increased to nearly 100%, and the accuracy of key audit targets has been tripled. The on-site customs have also shifted from the fatigue tactic of "people monitoring orders" to a new model of "system intelligent warning and manual precision strike".

At present, the system has been incorporated into the "Digital Customs Officer" system of Qingdao Customs, becoming the first AI application in China's customs to be put into practical use in the field of tax risk prevention and control.

From "1" to "N": replicable transfer of innovative methodology

The success of the first AI model not only solved a specific business problem, but also established a set of technical research methodology within the branch that emphasizes "party building leading, business driven, and rapid development". This methodology quickly demonstrated strong vitality in other innovative scenarios.

Faced with the dilemma of the implementation of the "inland waterway freight deduction" policy being hindered by the inability of enterprises to obtain segmented freight vouchers, the branch did not stay at the level of policy explanation, but launched technical research again.

Party members have developed an intelligent plugin embedded in the international trade "single window" system. This plugin can automatically identify eligible domestic branch line transportation segments, and based on the main domestic branch line freight market surveyed, automatically fill in deduction information, achieving "one click access" to complex preferential policies.

In 2025, this innovative combination of "policy+technology" will reduce taxes for enterprises in the jurisdiction by 1.1219 million yuan, a year-on-year increase of 6.8 times, and transform policy dividends into real money for enterprises.

Mechanism Escort: Building a "Firewall" and "Accelerator" for Continuous Innovation

Continuous technological innovation cannot be separated from a healthy mechanism soil. The Party branch deeply recognizes that innovation, especially technological exploration, inevitably comes with uncertainty and potential risks. To this end, the branch will deeply embed the "three item deliberation" system into the entire innovation process.

Before each innovation project is approved, it is necessary to first discuss integrity: comprehensively sort out the integrity risk points in technical decision-making power, data access power, and model interpretation power. During the project tackling, regular "business discussions" are held: in the form of technical democratic meetings, difficulties are collectively analyzed and explored through multiple paths. After the completion of the project, it is necessary to "discuss the team": evaluate the shaping of talent abilities through technological breakthroughs, and commend the pioneers who overcome difficulties.

This mechanism allows us to dare to innovate without blindly innovating, "said Jiang Shangjie, a propaganda committee member of the branch." It not only gives us the space to explore 'crossing the river by feeling the stones,' but also constantly reminds us of the bottom line of 'moving steadily and achieving long-term goals.'

Nowadays, the technological innovation of the branch has shifted from "tough battles" to "normalization".

Since conquering the first AI model, this branch has firmly embarked on a practical path of leading technological innovation with party building and empowering customs governance with technology. (Jiang Rong, Chen Cangqiu/text)