UIUC AI Teaching Assistant
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- The AI Teaching Assistant runs 11 separate models in parallel for text/image retrieval, generation, moderation, and ranking, achieving a median 2-second response time on HuggingFace
- The system draws on textbooks, lecture videos, and student QA forums, none of which are publicly distributed because the project was not granted those rights by the authors
- Creator kastan developed a novel semantic search retrieval approach during RLHF, using a comparison dataset produced by a hired team of five Electrical Engineering students and released freely on HuggingFace
- The project targets UIUC's ECE 120 (intro to Electrical Engineering) and is fully open source except for commercial textbooks, with Pinecone-based document store scripts included so users can plug in their own material
- Evaluation uses GPT-3 to judge model-generated answers against human-written ground truth, though the author flags the limitation that GPT-3 tends to rate its own outputs favorably — and suggests re-running with Cohere's models for comparison
Why it matters: For course-specific AI tutors, the project demonstrates a working 11-model orchestration pattern with 2-second median latency and a reproducible RLHF dataset pipeline. The author openly flags GPT-3 self-evaluation as a credibility gap, which is rare; ECE 120 students and instructors get a free, plug-and-play Gradio app, while the openly released RLHF QA comparisons lower the barrier for anyone building domain-specific tutoring systems.
