Recently, Unisound received a letter of commendation from China Railway Shanghai Design Institute Group Co., Ltd. (hereinafter referred to as “Shanghai Railway Design Institute”). In the letter, the Institute highly praised the Unisound team’s technical strength, project delivery, and service attitude in the “Large Model Cloud Service and Intelligent Agent Application Development Service Project”.

Figure 1: Letter of commendation from China Railway Shanghai Design Institute Group Co., Ltd.

Behind the letter of commendation lies the real-world practice of Unisound’s AI technology deeply embedded in business scenarioses.

As a large comprehensive research institution, the audit work of Shanghai Railway Design Institute has long faced three practical challenges: historical audit reports and institutional documents are scattered and fragmented, making it time-consuming and labor-intensive to find supporting references; document review relies on manual item-by-item comparison, which is inefficient, inconsistent in standards, and prone to omissions; and a large amount of audit experience is held by individuals, making it difficult to form sustainable, reusable organizational knowledge. At their core, these problems stem from unstructured documents that are hard to understand, domain knowledge that is hard to accumulate, and business experience that is hard to reuse.

To address these pain points, Unisound has followed a mainline of “holistic planning + technology R&D + knowledge engineering + delivery and operations”, integrating technical capabilities such as document parsing, knowledge graphs, RAG retrievals-augmented generation, large models, and intelligent agents, to build the “Shanghai Railway Smart Audit” intelligent audit computing platform for the Institute, bringing AI into the core audit workflow in four steps.

01 | Intelligent Parsing: Turning “Dormant Documents” into “Living Knowledge”

Audit materials are numerous, varied in format, and stored in a scattered manner — being able to “find” them does not mean being able to “use” them. Leveraging its intelligent document parsing capability, Unisound has performed unified parsing of the Institute’s audit reports, institutional documents, and project materials over the years, building an audit material repository, an institutional document repository, and an issue repository. To date, the platform has aggregated audit materials from 2017 to 2026, accumulating more than 2,000 institutional documents and more than 200 issue-list documents, turning scattered historical materials into truly searchable, linkable, and reusable knowledge assets.

02 | Precise Retrievals: Turning “Flipping Through Files” into “Finding Answers”

Faced with a vast number of institutional documents and historical cases, auditors in the past often had to repeatedly flip through and compare materials before finding the corresponding basis. Unisound has built hybrid retrievals and multi-level reranking capabilities tailored to audit scenarioses, combining keyword retrievals, semantic retrievals, and an audit knowledge graph to achieve deep linkage among institutional documents, review points, and historical issues, enabling AI to quickly find relevant supporting evidence.

03 | Verifiable Output: Making AI Audits “Accurate” and “Evidence-Backed”

Professional audit scenarioses require AI not only to give conclusions, but also to ensure that conclusions are evidence-based and the process is traceable. Through RAG, knowledge graphs, and output verification mechanisms, Unisound enables AI to analyze based on institutional documents, cases, and review points, and automatically associates each audit conclusion with the institutional document number, clause number, and original source, forming a complete chain of evidence. To date, the platform has developed 50 review points for outsourced contracts and 30 for research projects, with an accuracy rate of 80%–90% on related verification projects — AI not only answers, but can also provide the evidence.

04 | Long-Document Understanding: Keeping “Extra-Long Documents” Out of AI’s Blind Spots

Audit materials are lengthy and information-rich, and simple segmentation can easily break the context. Unisound adopts an approach of “semantic segmentation + causal/dependency chain restructuring”, preserving the relationships between content segments while decomposing long documents, so that AI understands both local information and the logic across the text, truly “comprehending” extra-long professional documents.

Notably, the collaboration goes beyond project delivery — the two sides have also joined forces to jointly tackle major research topics in the railway industry. Unisound’s large model cloud services and intelligent agent development capabilities are also moving from single-project delivery toward deeper empowerment of scientific research and business innovation.

From “finding materials” to “finding evidence”, from “manual verification” to “AI-assisted audit”, and then to turning historical experience into reusable digital knowledge — Unisound is bringing AI into the very core of the audit workflow.