TeachBack AI

Inspiration

Students often reread notes and feel that they understand a topic, but recognising information is different from being able to explain it clearly. I wanted to create a tool that tests active understanding instead of simply summarising information.

TeachBack AI was inspired by the teach-back/Feynman learning approach: explain what you know, identify what is missing, practise the weak area, and try again.

What it does

TeachBack AI turns study notes into an active learning loop.

Students add their notes, choose a concept, and explain it in their own words. The system analyses the explanation using semantic similarity, provides an understanding score, identifies important missing ideas, and gives a targeted diagnostic challenge based on the detected knowledge gap.

Students can then practise weaker concepts and track their progress over repeated attempts.

How we built it

TeachBack AI was built with Python and Streamlit.

The AI/NLP system uses Sentence-BERT (all-MiniLM-L6-v2) for semantic text representation, TF-IDF n-grams for concept extraction, and cosine similarity for comparing student explanations with relevant learning material.

I also implemented knowledge-gap detection, targeted diagnostic challenges, adaptive practice logic, progress tracking, and a TF-IDF fallback if the Sentence-BERT model is unavailable.

AI tools, including ChatGPT, were used as a development assistant for brainstorming, debugging, code refinement, and documentation. I tested, integrated, and refined the final application and learning workflow.

Challenges we ran into

The biggest challenge was making the feedback useful rather than simply producing a similarity score.

A student might understand the basic definition of a concept while still missing an important idea. I therefore refined the assessment process so that TeachBack AI identifies specific gaps and turns them into focused follow-up practice.

Another challenge was keeping the application reliable while using a machine-learning model, so I added a TF-IDF fallback if Sentence-BERT cannot be loaded.

Accomplishments that we're proud of

I am particularly proud of creating a complete adaptive learning loop rather than another note summariser.

TeachBack AI can move from a student's explanation to a measurable understanding score, identify what is missing, and target that specific weakness with another challenge.

For example, during testing an explanation received a 62% understanding score. TeachBack AI identified a missing idea, generated a targeted diagnostic challenge, and the improved response achieved a 91% score.

The application is also deployed online and works without requiring a paid AI API.

What we learned

Building TeachBack AI gave me practical experience with NLP, semantic embeddings, similarity scoring, adaptive learning logic, Streamlit deployment, debugging, and interface design.

Most importantly, I learned that AI in education does not have to simply generate answers for students. It can help students discover what they do and do not understand.

What's next for TeachBack AI

Next, I would like to add document and PDF uploads, richer personalised learning histories, additional question types, and evaluation with real student users.

A longer-term goal is to investigate whether repeated teach-back practice and targeted knowledge-gap feedback can produce measurable improvements in student learning.

Built With

• cosine
• machine-learning
• natural-language-processing
• python
• scikit-learn
• sentence-bert
• streamlit
• tf-idf