AI in Education: HSE University Case Studies

Alexey Masyutin24 May 2025

Problem

A university runs its own English certificate exam, and every candidate is scored on 10 criteria (5 for writing, 5 for speaking) by two professional linguists, plus a third when they disagree. This is heavy routine work for experts. Separately, the same university has more than 55 thousand students who ask the study office the same organizational questions, which costs about one person-month on the most typical question alone, even at two minutes per answer.

Example from the article: the HSE English test had 3,000+ candidates in 2023-2024 (slide 6 says 3,500+), and the linguists' criterion-level grades became the training set for an automatic grader.

Solution in brief

Treat expert graders as annotators and train a model to reproduce their grades: a fine-tuned small model for essays, and audio plus text embeddings with CatBoost for speech. For organizational questions, use a standard RAG assistant over 200+ regulatory documents. Check graders against a held-out set of expert-scored exams.

Questions this case answers
  • How to automate scoring of open-ended answers in text and audio so that the scores reproduce the consensus of expert raters
  • How to answer frequent questions from a large set of regulatory documents with a retrieval-augmented assistant
  • How to automate scoring of open-ended exam answers (essays and speech) so that it matches expert raters
  • How to build a RAG assistant over regulatory documents that answers student questions

Tools and methods

  • LoRA-adapted GigaChat — first essay grader, MAE about 0.7
  • DistilBERT with MAE + cross-entropy loss — final essay grader, 5 criterion grades
  • Whisper embeddings, text embeddings, CatBoost — speech grader
  • RAG: Tiktoken chunks, GigaChat embeddings, ChromaDB, GigaChat LLM — student office assistant
  • NER and embedding similarity — matching vacancy requirements to course content
Continue reading — it's free

Google or a one-time email code. No password.