Providing timely, personalised feedback in large-scale learning environments remains a persistent challenge. Generative artificial intelligence offers scalability, yet limited interpretability and concerns around ungrounded feedback constrain its use for formative assessment. This Work-in-Progress paper presents an early-stage hybrid pipeline that combines stylometric analysis with large language models (LLM) to support interpretable writing feedback at scale. Linguistic features are projected into low-dimensional, pedagogically interpretable indicators using Principal Component Analysis. Across two essay corpora comprising 2,828 student texts (PAN14 and ASAP), we observe stable patterns corresponding to two dominant dimensions of writing style: Syntactic Complexity and Lexical Richness. These dimensions are used to structure LLM-generated feedback and to support targeted diagnostic comments and cohort-level visualisation for instructor oversight in large classes. Ongoing work will examine how this approach informs instructional decision-making and student writing development in authentic learning contexts.
