AutoTA

• 

  AutoTA guides a student through the lesson with the objective of addressing students shortcomings and aiding in the learning process.

Inspiration

When teaching I try my best to help students fully understand every concept and topic, but it's tricky. The more I know, the more I underestimate pitfalls that can lead to confusion and frustration in a student. I think direct and hands on conversations are crucial for learning, but I don't always have the time each student requires.

I developed AutoTA to provide that hands on feedback; to allow students to test their understanding and explore information organically.

What it does

AutoTA ingests a transcript (or topic) and constructs a ReAct agent designed to communicate with a student. By using the encyclopedic knowledge of Gemini models, paired with tools designed to interact with the student in meaningful ways, an engaging and organic exploration of course material and concepts can be achieved.

How I built it

It started in Colab. I built a ReAct style agent and developed the core tools that produced compelling user experiences. I then built a front end UI using p5.js, and hosted it on Github pages.

Challenges I ran into

designing a UI to seamlessly integrate with a language model can be tricky. The response from the language model has to be correctly parsed, and the prompting needs to deftly balance user experience, flexibility, and robustness. I developed some re-prompting and error handling within AutoTA, but I think more advanced schemes (like a multi agent system, and more robust parsing and re-prompting) could result in an even better experience.

Also, naturally, with the challenges of building a UI and making things robust, I didn't have as much time to experiment with tools as I would have liked. I think AutoTA is great (I learned a surprising number of things from it) but I think more tools would allow for emergent interaction that would be even more compelling.

Accomplishments that I'm proud of

I think it looks good, feels great, and is genuinely useful. I'm very proud of AutoTA

What I learned

I learned much more about building agentic systems, and some of the realities of getting them to work within a larger context. I'm excited to continue learning more.

What's next for AutoTA

It would be great to get some feedback, maybe use a more robust front-end framework, allow for additional tools, and work with a diverse team to make AutoTA sing. I think AutoTA is special, and could be even more so.

Built With

• colab
• javascript
• p5.js
• react