Educational institutions increasingly collect large volumes of qualitative feedback through surveys, yet traditional qualitative analysis methods are often impractical due to resource constraints and reporting deadlines. This paper presents a methodological framework that leverages large language models (LLMs) to conduct qualitative analysis at scale. Our work addresses challenges at the National Institute of Education, Singapore, where we annually collect approximately 100 pages of open-ended responses from beginning teachers but lack the specialised expertise to analyse this data comprehensively. Rather than simply replicating traditional sentiment analysis approaches, we developed a framework that harnesses LLM affordances to transform both how qualitative analysis is conducted and how results are consumed. The framework employs aspect-based sentiment analysis principles guided by four key considerations: strategic framing to transform operational responses into strategic insights, contextual understanding to interpret intent, selective extraction to redefine neutrality, and completeness to capture emotional nuances. Our approach introduces “ambivalent” as a fourth sentiment polarity and incorporates empirically-derived sentiment qualities and themes rather than stakeholder-prescribed categories. Unlike traditional qualitative analysis that produces singular narratives, our framework generates multidimensional analytical outputs that enable stakeholders to explore findings non-linearly while maintaining analytical rigour. The methodology includes reliability assessment mechanisms to filter low-quality responses and strategic interpretation to align outputs with institutional decision-making needs. This work demonstrates how computational approaches can move beyond data treatment to enable genuine qualitative insights at scale, offering a pathway for educational institutions to transform overwhelming feedback into actionable intelligence that informs policy decisions and programme improvements.