Persona-Driven Feedback Agents for High-Stakes Documents: A Multi-Perspective Evaluation of Global Talent Visa Submissions

Adanyin, Anthonette Chidinma, Fagbohun, Oluwole, Akintola, Akinyemi S., Oyeladun, Oluwatosin, Akwaowo, Promise Udeme, Okere, Chinomso, Anyogo, Felicia Ene and Olajuwon, Olaonipekun Olaitan (2026) Persona-Driven Feedback Agents for High-Stakes Documents: A Multi-Perspective Evaluation of Global Talent Visa Submissions. In: 2026 International Conference on Advances in Artificial Intelligence and Machine Learning (AAIML). IEEE, pp. 367-372. ISBN 9798331568061 (In Press)

Full text not available from this repository.

Abstract

In this paper, we present a persona-driven multiagent AI framework for the automated evaluation and optimisation of UK Global Talent Visa (GTV) applications in the Digital Technology category. Inspired by persona-based reasoning and debate-oriented architectures, our approach assigns each agent a specialised persona, for example, GTV Application Assessor or Evidence Quality Examiner, to analyse a distinct component of the application. The architecture is orchestrated using CrewAI and powered by the OpenAI o3-mini reasoning model, enabling structured and parallel evaluation of applicant documents through coordinated agent collaboration. Each agent returns Pydantic-validated JSON outputs that guarantee type safety, interpretability, and reproducibility across the assessment pipeline. The system consolidates all applicant materials, including personal statements, CVs, recommendation letters, and evidence documents, and evaluates them against four assessment pillars: Innovation, Impact, Recognition, and Leadership, as formalised by Tech Nation. Specialised agents generate quantitative readiness scores, qualitative feedback, and evidence-based recommendations that classify applications as Strong (≥85%), Viable (60−84%), or Non-viable (<60%). Empirical evaluation shows that the complete multi-agent workflow achieves full application assessment in approximately three to four minutes with a 100% schema validation success rate. Persona-driven specialisation enhances interpretive depth and feedback consistency compared with monolithic LLM approaches. The integration of structured schema validation, multi-agent orchestration, and empirically grounded scoring heuristics establishes a robust foundation for large-scale, explainable AI studies in high-stakes immigration and policy evaluation contexts.

Item Type:	Book Section
Uncontrolled Keywords:	Multi-agent systems; Large language models; Persona-driven agents; Structured output validation; CrewAI; OpenAI o3-mini; Pydantic; Immigration policy analytics; Explainable AI; Global Talent Visa (GTV)
Subjects:	Q Science > Q Science (General) > Q336 Artificial intelligence
Divisions:	Schools and Research Institutes > School of Business, Computing and Social Sciences
Depositing User:	Kamila Niekoraniec
Date Deposited:	12 Aug 2026 13:06
Last Modified:	12 Aug 2026 13:06
URI:	https://eprints.glos.ac.uk/id/eprint/16366

University Staff: Request a correction | Repository Editors: Update this record