At the 8th China Fintech Forum, held on September 9th in Beijing under the theme "Technology Empowering - Digital Innovation and Application in the Financial Industry," Che Zhongliang, President of the Financial Business Group at Isoftstone Information Technology (Group) Co., Ltd., delivered a keynote speech on AI software engineering practices.
Che began by explaining the dual meaning embedded in his presentation title, "AI Software Engineering Practice." The first part reflects his company's core mission: serving financial institutions through software engineering, a field that has become the primary arena for AI-driven transformation and disruption. The second part, "practice," encompasses both the challenges encountered along the way and the pragmatic, hands-on approach his team has adopted in tackling these issues.
Reflecting on the past year, Che noted that at the same venue in 2023, he had discussed how AI was reshaping software engineering paradigms. Over the subsequent twelve months, as his team navigated this transition, they encountered a critical question: what exactly constitutes the new paradigm when AI takes charge of driving processes forward? The previous framework seemed to have disappeared, replaced by uncertainty about how this new approach could gain customer acceptance and how they could demonstrate its validity through both internal validation and external recognition.
Drawing on real-world experiences, Che highlighted the growing prevalence of VibeCoding, a powerful approach where AI handles a vast array of tasks with remarkable efficiency. Simultaneously, many enterprises continue to rely on Copilot-style assistance. The appeal of VibeCoding lies in AI's tireless dedication, its willingness to accept endless revisions, and its capacity for proactive thinking and iterative self-validation. However, this raises a crucial question: who assumes responsibility for the outcomes? Echoing concerns raised by a previous speaker, Che emphasized the challenges of observability, explainability, evaluability, and intervention in AI-driven processes.
Che drew an analogy to illustrate the unpredictability of AI reasoning. Just as rural elders can set traps along a rabbit's familiar path near its burrow, large language models exhibit significant variability in their reasoning processes—perhaps even more so than biological systems. This variability makes it difficult to explain many aspects of execution, particularly during VibeCoding or when handling long-range tasks where pausing for intervention mid-process proves challenging.
To address these issues, Che's team developed a strategy of segmenting complex software engineering processes into manageable phases. Drawing on the simple question of how to fit an elephant into a refrigerator—by dividing it into steps—they dismantled robust engineering workflows into discrete sections. Each segment incorporates standardized methods, protocols, embedded AI agents and assistants, and quality gates with defined thresholds. This structure allows AI to operate freely within trusted parameters while enabling human intervention at critical junctures, whether to provide guidance, take over specific components, or assume full control when necessary.
This approach transforms what might otherwise be a "black box" task delegated to VibeCoding into a "white box" process with full visibility, measurement capabilities, and evaluative frameworks. It revisits traditional software engineering elements—requirements, design, testing, test cases, test outlines, structure, pages, and transitions—to determine whether they remain relevant and how AI can generate these deliverables while allowing human review, quality inspection, and document-driven progression.
Che cited inspiration from a CCTV program "Great National Infrastructure," which showcased projects like the custom-built climbing machine for Shanghai Tower and the specialized floating vessel used by CCCC First Highway Engineering Bureau to transport massive tunnel segments for the Shenzhen-Zhongshan Bridge. Similarly, his team recognized the need to construct a tailored engineering system for their clients—one that is segmented, standardized, and customized with specific management protocols governing every phase.
This framework bridges the gap between VibeCoding and Copilot, establishing a complete chain that spans reverse engineering for requirement clarification, specification design, assembly, production, and acceptance. Throughout this pipeline, traditional software engineering documents and artifacts—now produced with AI support—remain open to human intervention, iteration, and handoff back to AI for further processing. This enables continuous execution across long-range tasks, batch iterations, and version upgrades.
The architecture creates a deterministic path for AI enablement while providing humans with a visible channel for observation, evaluation, modification, and even full takeover. From project initiation through definition, construction, and acceptance, a human-machine collaboration mechanism ensures AI handles the majority of work while maintaining human visibility. The guidance, opinions, reviews, and decisions provided by humans direct AI along a relatively stable path, allowing information to flow through a well-defined chain.
Che emphasized the multiplicity of "abilities" that matter in this context: observability, constraint, intervention, verification, and auditability. The team's goal is to ensure AI outputs are subject to these controls, while the toolchain itself demonstrates capabilities in definition, configuration, control, and standardization, all while enabling execution paths to be reproducible, restartable, and resumable from any interruption point.
In closing, Che acknowledged the inherent difficulty of this endeavor—fitting a highly flexible, seemingly omnipotent AI into a moldable, standardized pipeline requires substantial ongoing effort. He expressed gratitude for the audience's attention and concluded his presentation with a commitment to continuing this challenging but necessary exploration.