The Summer Undergraduate Research Fellowship (SURF) program at XJTLU provides invaluable research experience for STEM students, but faces significant scalability challenges due to repetitive supervisory tasks and students' difficulty navigating complex research onboarding processes. This article presents SurfMate, an AI agent developed on the XIPU AI Agent Platform to address these dual challenges. SurfMate employs a phase-specific, workflow-driven architecture that mirrors the SURF program lifecycle, supporting students through interview preparation, program onboarding, and project execution. By combining a fixed Q&A Library for consistent information delivery with a dynamic Knowledge Base enhanced by real-time search plugins, SurfMate provides comprehensive support while automating administrative tasks such as interview scheduling and progress tracking. The agent also features an LLM-powered CV tailoring workflow that guides students through structured questions to generate program-specific applications. Expected outcomes include a reduction in supervisory workload, enhanced student self-directed learning, and improved matching quality between students and programs. This study demonstrates how strategically designed AI agents can transform undergraduate research supervision into a more scalable, efficient, and engaging educational experience.

AI Agent; Undergraduate Research; STEM Education; SURF Program; XIPU AI Platform

 

Undergraduate research programs have become cornerstone experiences in research-led universities, providing students with authentic opportunities to engage in scientific inquiry. The Summer Undergraduate Research Fellowship (SURF) program at Xi'an Jiaotong-Liverpool University (XJTLU) exemplifies this commitment, offering STEM students invaluable experiential learning. However, the effectiveness of such programs is fundamentally constrained by a scalability issue in mentorship. Supervisors make significant time commitments to repetitive administrative and foundational explanatory tasks, particularly during initial program screening and onboarding phases. This includes repeatedly explaining program details to numerous applicants and guiding new members through core literature. This high volume of repetitive inquiries creates a substantial bottleneck, limiting the number of students a supervisor can effectively mentor and diverting time from high-value, personalized scientific guidance. Simultaneously, students face complementary challenges. With limited initial information, they may struggle to assess their genuine interest or fit for a program before interviews, leading to potential mismatches. Furthermore, the transition into specialized research can be daunting, as students must rapidly assimilate background knowledge from complex academic literature, often without immediate, tailored support. This dual challenge represents a classic educational scalability problem, similar to Bloom's "two-sigma problem" where personalized instruction at scale remains elusive [1].

Recent advances in generative AI and AI agent technologies offer promising solutions to such scalability challenges [2]. AI agents, autonomous or semi-autonomous systems capable of performing tasks that typically require human intelligence, are increasingly integrated into educational settings to address challenges in scalability, accessibility, and individualized learning support. The XIPU AI Agent Platform at XJTLU provides an institutional foundation for developing tailored AI solutions. Building on the university's Education + AI Strategic Framework 2025-2028, which emphasizes symbiotic relationships between human and digital intelligence, educators have begun deploying AI agents for a range of educational purposes. Examples include module management agents that provide instant, document-verified responses to student queries [3], specialized agents for EAP exam preparation [4], and multiple agents within individual courses to address distinct pedagogical needs [5].

This article presents SurfMate, an AI agent designed specifically to enhance the SURF program experience. SurfMate serves as an intelligent assistant throughout the research lifecycle, from initial applicant screening through project execution, automating routine tasks while providing on-demand support that complements human mentorship. This case study demonstrates how strategically designed AI agents can transform undergraduate research supervision into a more scalable, efficient, and engaging educational experience.

 

SurfMate's architecture is structured around a "Modular Agent + Workflow Orchestration" logic that directly mirrors the three-phase workflow of the SURF program: Interview Preparation Phase, Program Onboarding Phase, and Program Execution Phase. This phase-specific design ensures that the AI agent provides targeted support aligned with users' current needs, moving beyond generic chatbots to deliver contextual assistance throughout the research lifecycle.

The agent initiates interactions with a welcoming conversation opener (Figure 1) and provides six quick commands that guide users through phase-appropriate tasks. This workflow-driven approach, orchestrated by the platform's Workflow Orchestration Engine, creates a seamless user experience in which triggering a stage-specific task automatically initiates a chain of actions, calling the necessary tools (Q&A Library, LLM, or plugins) in a predefined sequence. Such orchestration ensures that the AI agent consistently provides appropriate support while maintaining the human touch essential to mentorship.

 

 

SurfMate employs a "Fixed Q&A + Dynamic Expansion" methodology for knowledge base construction, ensuring both consistency and comprehensiveness in information delivery.

 

For standardized, process-oriented information, such as program background, literature guidelines, and poster requirements, the platform's native Q&A Library functionality is used to create a repository of fixed, accurate responses. This ensures rapid, consistent information delivery while maintaining accuracy of critical program details that should not vary. The Q&A Library was chosen for this purpose rather than a general knowledge base because it prevents unintended modifications when retrieving factual data and automatically generates direct links to resources when provided with exact file names.

 

For domain-specific terminology and concepts, a complementary approach is employed. Core terms and their definitions extracted from key literature are first uploaded to the platform's Knowledge Base. This base is then integrated with the "BochaAISearch Plugin," establishing a two-tier retrieval system: the agent prioritizes searching the internal knowledge base, and only for uncovered queries does it leverage the plugin for real-time, supplementary web searches. This hybrid approach guarantees comprehensive coverage while maintaining accuracy for known information.

 

 

The AI agent provides distinct functionalities across the three SURF phases:

 

The Program Information Query function utilizes the Q&A Library capability to provide instant, standardized program descriptions, complemented by direct links to software tutorials for beginners to facilitate rapid skill acquisition (Figure 2).

 

The CV Tailoring feature leverages a custom LLM-powered workflow to intelligently parse user input and generate a structured, program-tailored CV (Figure 3). To prevent students from including irrelevant experiences, this feature guides them through detailed questions about their research background. For privacy protection, students are only required to provide a nickname during this initial stage, excluding sensitive information such as their real name, student ID, or phone number.

 

The Interview Slot Management function integrates the "Feishu Plugin" through a dedicated workflow to automatically compile user-provided availability into a centralized scheduling sheet, streamlining interview coordination (Figure 4).

 

Program Onboarding Phase:
The Literature Guidance function reuses the Q&A Library to deliver fixed guidelines for literature review (Figure 5). Key review papers recommended for beginners in the field are pre-uploaded to the Knowledge Base for easy access.

 

The Terminology Explanation function activates the core "Knowledge Base + BochaAISearch Plugin" integration, executing the prioritized retrieval logic to provide accurate and context-aware definitions of specialized concepts (Figure 6).

 

Program Execution Phase:
The Poster Guideline Query function deploys the Q&A Library to supply fixed poster design requirements, ensuring consistency in final deliverables (Figure 7).

 

The Task and Progress Management function relies on the Spreadsheet Plugin workflow to automatically update and organize team tasks and progress data into a shared table using the Feishu Plugin, reducing administrative overhead for both students and supervisors (Figure 8).

 

 

SurfMate serves three primary user groups:
1. SURF Applicants: Undergraduate students seeking to join SURF programs
2. SURF Members: Students who have been selected for SURF programs
3. SURF Supervisors: Faculty members overseeing SURF programs
The interaction flow follows the SURF program lifecycle, with the AI agent providing increasingly specialized support as users progress through the stages.

 

A student applicant discovers a SURF program through campus announcements and wants to assess their fit before applying. The supervisor provides the student with a link to access the AI agent via email. The agent provides a comprehensive program overview using the Q&A Library, along with links to beginner-friendly tutorials for required software skills. The student then uses the "Tailor Your CV" feature, which guides them through detailed questions about their research background and generates a program-specific CV that helps supervisors quickly identify qualified candidates.

If, after this preliminary exploration, the student remains interested, they proceed to schedule a face-to-face interview via the integrated Feishu Plugin. During the scheduled interview, the supervisor uses the student's pre-prepared, program-tailored CV as a basis for discussion, assessing the student's understanding of the program and their autonomous learning capacity based on their engagement with the AI-provided materials.

This efficient pre-screening process significantly reduces supervisor time investment while ensuring better-matched applicant-program pairs.

A newly selected SURF member needs to rapidly acquire program-specific knowledge and background. The system recommends curated review papers pre-uploaded to the Knowledge Base through the "Decode Key Literature" quick command in the AI agent, delivering a structured literature guide via the Q&A Library.

When the member encounters specialized concepts, they query the Knowledge Base with "BoCha AISearch Plugin" fallback, receiving concise explanations and summaries of complex terminology. This accelerates the learning curve, substantially reduces basic queries to the supervisor, and builds stronger foundational knowledge.

SURF team members update progress metrics using Feishu Plugin workflows, while the system tracks milestones and deadlines automatically. Supervisors monitor overall progress through consolidated dashboard updates. The agent also provides detailed poster and presentation guidelines via the Q&A Library, ensuring that final deliverables meet program standards.

This approach enhances program coordination and minimizes administrative overhead, allowing both students and supervisors to focus on research quality rather than administrative logistics.

 

 

This study conducted a systematic analysis of two SurfMate agents with different topics (seven agents have been created in total for different SURF topics), based on 63 user conversation records collected from June to August 2026. Synthesizing the data, a balanced evaluation can be made from both strengths and limitations.

In terms of strengths, SurfMate demonstrated substantial practical value in core academic support tasks. Data revealed that 82.9% of user inquiries concentrated on "project information inquiries" and "literature recommendations" (see Figure 9), indicating that the agent effectively addressed students' two primary needs during the research onboarding phase: understanding project scope and accessing relevant literature.

 

The knowledge base achieved a 100% hit rate across all documented queries, confirming adequate content coverage and stable retrieval performance. At the interaction level, the average conversation length reached 3.94 rounds, with 65% of dialogues extending beyond a single turn (see Figure 10). Notably, 30% of conversations exceeded 4 rounds, suggesting that a subset of students moved beyond simple "ask-and-leave" patterns and engaged with SurfMate as a tool for deeper research exploration. Furthermore, users predominantly triggered interactions through predefined functional phrases (e.g., "Decode Key Literature"), indicating that the interface design was intuitive and lowered the barrier to usage.

 

In terms of limitations, the data analysis revealed several areas requiring improvement. First, functional usage was highly imbalanced. Only 17.1% of inquiries fell into other categories such as "process inquiries," "daily conversation," or "poster guidelines," suggesting that the visibility or utility of these features remained insufficient, leaving them underutilized. Second, although multi-turn dialogues were observed, the platform has not yet enabled long-term memory functionality, preventing the agent from recognizing returning users or tracking historical context across sessions, thus limiting its capacity for personalized assistance. Third, monthly activity trends showed a peak in June, followed by a gradual decline in July and August (see Figure 11), suggesting insufficient sustained promotion during the later stages of the project, with users exhibiting a "leave after use" pattern rather than retaining engagement.

 

 

The integration of SurfMate into the SURF program represents a strategic approach to addressing the scalability challenges inherent in undergraduate research mentorship. By automating repetitive administrative tasks and providing phase-specific support for students, the AI agent enables supervisors to focus on high-value mentorship while empowering students to take ownership of their research journey. SurfMate's hybrid knowledge architecture—combining a fixed Q&A Library with a dynamic Knowledge Base enhanced by real-time search capabilities—ensures accurate, comprehensive support across diverse queries. Its workflow-driven, phase-specific design creates a more efficient, engaging, and scalable research experience for all involved. This case study demonstrates how thoughtfully designed AI agents can transform educational experiences while maintaining the human touch essential to effective mentorship. As institutions increasingly seek to scale high-impact educational practices, AI agents like SurfMate offer a promising path toward more accessible, equitable, and effective undergraduate research experiences.

 

 

[1] Bloom, B. S. (1984). The 2 sigma problem: The search for methods of group instruction as effective as one-to-one tutoring. Educational Researcher, 13(6), 4–16; Sarkar, T. (2025). The evolution of mastery learning: Challenges, technologies, and the AI journey leading to the intelligent tutoring system DARTS. Issues in Information Systems, 26(4), 45.
[2] Owan, V. J. (2026). Artificial intelligence as a pedagogical companion for STEM research learning in higher education. Discover Education.
[3] Jarvest, O., Li, K., Song, J., & Zhou, M. (2025). Module Management AI Agents: A Targeted Approach to Study Support in Higher Education. CEIE Articles, Learning Mall.
[4] Dai, Y., Zhou, J., Zhou, Y., Zhang, Q., & Xia, L. (2025). AI Agents as Scaffolds: Supporting ZPD and Self-regulated Learning in Exam Preparation. CEIE Articles, Learning Mall.
[5] Sun, Y., Cui, W., & Wu, Y. (2025). Applying Multiple AI Agents in One Course: A Case Study of the PGcert 404 Module. CEIE Articles, Learning Mall.