TAI Logo

Disclaimer: We were genuinely so close. We only needed to display everything on a streamlit app. Otherwise, all the functionality is there.

TAI – Tutoring AI for Indigenous Empowerment

Inspiration

Millions of Indigenous students across the world face a heartbreaking trade-off:
access to formal education often comes at the cost of losing their language and identity.

This vision led to TAI: a tutoring platform that uses Generative AI to deliver inclusive, culturally respectful education, in native languages, at scale.

What it does

TAI is an AI-powered tutoring platform that:

Delivers personalized educational lessons in Indigenous languages
Allows students to interact in real-time with a Gemini-powered chatbot tutor
Synthesizes curriculum-aligned content as it isn’t readily available in native languages
Generates adaptive quizzes based on lesson key concepts, to assess learning

The platform is designed to preserve identity while expanding access — making education more inclusive, equitable, and empowering.

How we built it

We combined multiple techniques and tools:

Frontend: HTML/CSS web interface for language and subject selection, chatbot, and lessons
Backend: Gemini API (Google Cloud) powers the multilingual chatbot and generates lessons and quizzes content

We also developed a deep learning research prototype:

Fine-tuned Google’s mT5 model using AmericasNLI — a dataset of Indigenous languages from the Americas
Focused on three Mexican languages: Rarámuri, Otomí, and Wixárika (Huichol)
Built a fine-tuning pipeline that can grow as more user-generated data becomes available

Challenges we ran into

Lack of high-quality datasets in Indigenous languages limited fine-tuning performance
Gemini prompt engineering required iterations to get high-quality, culturally respectful output
Creating a seamless interface for low-resource environments posed design constraints
Ensuring the AI-generated content remained accurate, culturally appropriate, and pedagogically sound

Accomplishments that we're proud of

Successfully integrated Gemini for multilingual tutoring in a real-time interface
Built a working quiz generation system based on lesson content
Developed a fine-tuning pipeline for low-resource languages, showing early promise
Proved that AI can be a bridge, not a barrier, for educational equity

What we learned

How to apply Generative AI to create localized educational experiences
Techniques for fine-tuning multilingual transformers (mT5) on low-resource language datasets
The importance of inclusive design and the challenges of building for underrepresented communities
That even with limited data, meaningful tools can be built to empower marginalized voices

What’s next

Expand language support by growing and curating datasets in collaboration with communities
Add voice input/output for oral-first languages or low-literacy users
Enable offline access for low-connectivity regions
Partner with local educators and Indigenous organizations to run pilot programs
Train more domain-specific fine-tuned Gemini prompts for STEM subjects

Reference

[1]: Ebrahimi, A., Mager, M., Oncevay, A., Chaudhary, V., Chiruzzo, L., Fan, A., Ortega, J., Ramos, R., Rios, A., Ruiz, I. V. M., Giménez-Lugo, G., Mager, E., Neubig, G., Palmer, A., Coto-Solano, R., Vu, T., & Kann, K. (2022). AmericasNLI: Evaluating zero-shot natural language understanding of pretrained multilingual models in truly low-resource languages. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 6279–6299. https://doi.org/10.18653/v1/2022.acl-long.435