Xiamen Metro Tech used Qoder to solidify its R&D process into standard workflows, cutting the average coding delivery cycle of a single feature module by about 30%.
Back
When AI enters software development, the easiest wins are local: code completion, content generation, and similar gains. The hard part is turning individual usage into stable, repeatable team delivery capability. If requirements, standards, reviews, and experience capture stay the way they were, AI only speeds up isolated steps and cannot change overall R&D efficiency.
For Xiamen Metro Innovation Technology Co., Ltd. (hereafter "Xiamen Metro Tech"), this problem is especially concrete. The company takes on complex projects with a team that is not large, and those projects involve cross-discipline collaboration, long-term iteration, and personnel rotation. Unifying engineering standards, shortening delivery cycles, and keeping project experience inside the organization matter more than simply typing code faster.
Every AI initiative has to solve the problem at hand and also leave behind reusable processes, knowledge, and engineering capability.
Based on that judgment, the company made software development the starting point for building AI capability. R&D runs through requirement analysis, design, development, review, and delivery; it is the production line every digital project passes through. Transform the way software is built first, and the resulting standards, processes, and knowledge can be reused again and again in later projects.
Since last November, Xiamen Metro Tech has adopted Qoder as its core R&D tool. The focus of the exploration is no longer "how much code AI can write," but "how people and AI complete a project together."
With that, Qoder moved from being a personal coding assistant into the team's R&D process, and the company's use of AI shifted from local efficiency gains to a restructuring of the development model.
The company set "small team + AI Agent" as its basic delivery unit. Team members are responsible for understanding the business, defining boundaries, making key judgment calls, and owning the outcome. AI takes on high-frequency work such as organizing information, generating specifications, repetitive coding, basic testing, and standards checking.
This model does not use AI to replace a role. It redistributes tasks between people and AI across the development process. People concentrate on business understanding, solution design, and quality control, while AI accelerates execution and keeps the process consistent.
Team members are also no longer each responsible for a single step in an assembly line. They can enter a project at the requirement stage and stay through design, development, verification, and delivery. That is how a small team gains the ability to own a complete closed loop.
To keep the new model from depending on individual habits, the company used Qoder to distill the development process into four stages — specification, build, gate, and delivery — and packaged the practices that recur in each stage into standard Skills.
Five Skills connect in sequence along the development process, each with a clear job:
●
The shift in the development model shows up first in delivery efficiency: the average coding delivery cycle of a single feature module is about 30% shorter, with repetitive work such as CRUD, interface integration, and unit testing increasingly handled by AI.
The deeper change is in how the team collaborates. Members can carry one requirement through its full lifecycle, which reduces the repeated handoffs, explanations, and waiting between steps. The organization starts shifting from division by role to ownership of closed-loop tasks.
The knowledge-capture mechanism also improved continuity on existing projects. With repository documentation, engineering standards, and accumulated project experience, onboarding time for new members went from "counted in weeks" to "counted in days," and personnel changes have less impact on project continuity.
These results show that the value AI brings has already gone beyond individual coding speed: standards are more unified, collaboration chains are shorter, and knowledge is easier to reuse. That is how a small team gains more stable project delivery capability.
Once the model proved itself inside the R&D line, the company applied it to complex business projects to test whether standard workflows and knowledge capture could hold up in real production scenarios.
The intelligent metro maintenance management system is the representative validation of this development model on a complex project.
Equipment maintenance is a critical part of metro operations, and its labor cost accounts for roughly half of total operational labor cost. The project had long faced paper-based procedures, reliance on individual experience, offline coordination, manual data entry, and data silos. It required the development team to understand complex business rules and required the system to handle multiple lines, multiple disciplines, and large-scale data.
The project did not start from algorithms and models. The development and business teams first unified spatial naming across the entire line network level by level, covering both trackside and non-trackside areas, governed equipment fields with templates under the principle of "unified framework, discipline autonomy," and bound every piece of equipment precisely to a specific location, establishing a unified data foundation for later system development.
Once data definitions were unified, the system followed digitalized procedures and scheduling rules, combining equipment procedures, staff skills, materials information, work windows, and operational constraints to automatically generate more than 100,000 maintenance plan work orders across multiple lines and disciplines, improving the overall efficiency of plan preparation, resource allocation, and work execution.
Qoder served as the R&D tool throughout: Vibe Coding accelerated development of key modules such as intelligent scheduling, conflict detection, and automatic generation of work orders and plans, while field definitions, spatial naming, and work standards were continuously captured as project assets, helping the team keep engineering conventions consistent over a long build cycle.
The project validated the practical value of the new development model: facing a complex system with tens of millions of data records and interwoven disciplines, a small team could still advance key module construction through standard processes, AI collaboration, and knowledge capture, while lowering the cost of personnel rotation and onboarding.
Looking back at the whole practice, Xiamen Metro Tech did not simply hand its development team one more AI tool. It reorganized the development process around "small team + AI Agent": first clarify the division of labor between people and AI, then solidify specification, build, gate, and delivery into workflows, and finally keep capturing project knowledge.
What Qoder provides in this model is connection and consolidation. It brings rules and experience scattered across different members and different stages into one development process, so the efficiency of AI can be shared by the team and project outcomes can be reused in later work.
The core of this model is not how much faster one round of development goes, but the gradual formation of stable organization-level R&D capability: requirements can be expressed accurately, engineering standards can be enforced continuously, quality issues can be caught early, and delivery experience can keep accumulating.
When tool usage is solidified into workflows and workflows are captured as organizational capability, AI truly moves from assisting with code to driving R&D productivity. And every project delivery becomes the starting point for the next round of development and innovation.
Change the Way You Build Software, Not Just Add One More AI Tool
When AI enters software development, the easiest wins are local: code completion, content generation, and similar gains. The hard part is turning individual usage into stable, repeatable team delivery capability. If requirements, standards, reviews, and experience capture stay the way they were, AI only speeds up isolated steps and cannot change overall R&D efficiency.
For Xiamen Metro Innovation Technology Co., Ltd. (hereafter "Xiamen Metro Tech"), this problem is especially concrete. The company takes on complex projects with a team that is not large, and those projects involve cross-discipline collaboration, long-term iteration, and personnel rotation. Unifying engineering standards, shortening delivery cycles, and keeping project experience inside the organization matter more than simply typing code faster.
Every AI initiative has to solve the problem at hand and also leave behind reusable processes, knowledge, and engineering capability.
Based on that judgment, the company made software development the starting point for building AI capability. R&D runs through requirement analysis, design, development, review, and delivery; it is the production line every digital project passes through. Transform the way software is built first, and the resulting standards, processes, and knowledge can be reused again and again in later projects.
Since last November, Xiamen Metro Tech has adopted Qoder as its core R&D tool. The focus of the exploration is no longer "how much code AI can write," but "how people and AI complete a project together."
With that, Qoder moved from being a personal coding assistant into the team's R&D process, and the company's use of AI shifted from local efficiency gains to a restructuring of the development model.
The Core Model: "Small Team + AI Agent"
The company set "small team + AI Agent" as its basic delivery unit. Team members are responsible for understanding the business, defining boundaries, making key judgment calls, and owning the outcome. AI takes on high-frequency work such as organizing information, generating specifications, repetitive coding, basic testing, and standards checking.
This model does not use AI to replace a role. It redistributes tasks between people and AI across the development process. People concentrate on business understanding, solution design, and quality control, while AI accelerates execution and keeps the process consistent.
Team members are also no longer each responsible for a single step in an assembly line. They can enter a project at the requirement stage and stay through design, development, verification, and delivery. That is how a small team gains the ability to own a complete closed loop.
Solidifying the R&D Process into Standard Workflows
To keep the new model from depending on individual habits, the company used Qoder to distill the development process into four stages — specification, build, gate, and delivery — and packaged the practices that recur in each stage into standard Skills.
Five Skills connect in sequence along the development process, each with a clear job:
● init-proj-rules: when spinning up a new project, first commit engineering standards and the design baseline into the repository;
● write-spec: write the requirement specification clearly, giving AI an unambiguous construction blueprint;
● start-dev: move into coding, where AI takes on high-frequency work such as CRUD, interface integration, and unit tests;
● review-mr: the pre-merge review gate, where AI checks standards and common issues;
● summarize-learnings: after delivery, feed effective practices and problem-handling experience back into the team knowledge base.
RepoWiki goes a step further and turns the code repository into readable documentation, continuously recording project structure, key modules, and engineering conventions, which lowers the risk of knowledge sitting with only a few people.
At this point, what Qoder connects is not just the coding step, but the full arc of a project: establishing rules, expressing requirements, executing development, checking quality, and capturing knowledge. R&D experience starts turning from individual practice into workflows the team can invoke repeatedly.