Project Background

In collaboration with the software team of a telecommunications equipment platform, we launched a pilot project focused on enhancing software R&D efficiency using LLMs. This initiative specifically addresses the client's key challenges, including a large proprietary codebase, strict internal coding standards, complex software interfaces and dependency relationships, as well as a high entry barrier due to specialized knowledge. To tackle these issues, we conducted target development optimization of models, training corpora, and associated tools, significantly improving large model performance in enhancing R&D productivity for complex system software.

Consultation Process

We first analyzed the client's unique business operations and software characteristics to assess their software delivery pipeline and key roles. Based on this analysis, we formulated practical goals for improving AI-assisted R&D efficiency and proposed a detailed implementation plan. Specifically, we assisted the client with: the selection and evaluation of foundational AI models; designing model fine-tuning strategies; developing plans for screening and optimizing code corpora, creating strategies for Prompts and Supervised Fine-Tuning (SFT); building a Retrieval-Augmented Generation (RAG) system for R&D knowledge and code APIs; optimizing tool agents; and, stablishing mechanisms for testing, evaluation, and feedback. Through these comprehensive measures, we assisted the client in enhancing performance and adoption of AI models in the development of complex software.

Collaborative Achievements

We assisted the client in establishing realistic goals and a strategic plan tailored to their unique industry and operational context. Furthermore, we helped enhance their capabilities for the continuous development, testing, and optimization of models and tools within their AI-assisted R&D framework. As a result, the client achieved a significant increase in the acceptance rate of AI-generated business code and test case code.

Project Background

Operating within their corporate intranet and constrained by limited computing resources, the client could only pilot small-scale LLMs to improve software R&D efficiency. The primary challenge was that their codebase contains numerous custom libraries and specific coding standards, making it difficult for general-purpose LLMs to accurately comprehend the context and generate compliant code.

Consultation Process

To address the client's intranet environment and computational constraints, we began by systematically defining AI application paradigms for software development and establishing a comprehensive conceptual framework. By optimizing the generation process and specifying key steps, we significantly enhanced the capabilities of their parameter-limited language models. To achieve this, we developed a suite of solutions: a RAG retrieval system for historical code, enabling the model to accurately understand the client's unique codebase; an external toolchain to automatically detect and correct AI-generated code, ensuring compliance and quality; an analysis tool for existing code that identifies patterns and rules to guide the generation process; a code refactoring tool to automatically modernize and optimize legacy code. Throughout the entire process, we worked in close collaboration with the client, continuously refining the solution, and creating a complete methodology and toolchain for large model assisted software development within a corporate intranet environment.

Collaborative Achievements

The project delivered measurable improvements, increasing the acceptance rate of AI-generated code from 30% to over 50%. Additionally, the generated code's stability was also increased by 80%, reducing the number of unintended modifications to existing projects. This led to a significant boost in software R&D efficiency, even under the client's resource constraints. The most significant gains were in the areas of standard component development and unit test generation. Ultimately, we established a sustainable, continuously optimizable workflow for AI-assisted development, setting the groundwork for future AI-empowered initiatives.