Stan: An LLM-based thermodynamics course assistant
A new AI assistant runs entirely on local hardware, offering privacy and cost control for university courses.
Researchers Eric M. Furst and Vasudevan Venkateshwaran have introduced Stan, a novel AI assistant designed for a chemical engineering thermodynamics course that uniquely supports both students and instructors from a single data foundation. Unlike typical educational AI focused solely on students, Stan processes lecture transcripts and a structured textbook index through a locally-run retrieval-augmented generation (RAG) pipeline. For students, it answers queries with specific page references. For instructors, it provides semester-scale analytics, identifying student questions, moments of confusion, and cataloging teaching analogies, creating a searchable record for course improvement.
The system is built entirely on open-weight models—Whisper large-v3 for transcription and Llama 3.1 8B for processing—running on local hardware to ensure data privacy, predictable costs, and independence from third-party APIs. The paper details practical challenges in deploying these 7–8 billion parameter models for long-context tasks, including context truncation and schema drift, along with the technical mitigations developed. This dual-role, self-contained architecture presents a scalable blueprint for deploying private, reproducible AI tools in specialized academic and professional training environments.
- Dual-role AI assistant built on a shared RAG pipeline using lecture transcripts and a textbook index.
- Runs entirely locally using open models (Whisper large-v3, Llama 3.1 8B) for full data privacy and no API costs.
- Provides instructors with analytics on student confusion and teaching patterns from a semester's worth of lectures.
Why It Matters
Offers a blueprint for private, cost-controlled AI deployment in education and specialized training, shifting focus to instructor support.