Learn Mentor

Inspiration

Personalised 1:1 tutorials have been proven to significantly enhance learning outcomes. However, the high cost makes them inaccessible to many. Recent advancements in AI, particularly with large language models (LLMs), offer a promising alternative—leveraging technology to provide personalised education.

What it Does

Learn Mentor is a simple app designed with various modes to assist students in specific tasks. For instance, it can help students conduct a pre-mortem analysis to mitigate project failures. Another example is providing constructive feedback and suggestions for improving essays, without actually writing the essays for the students.

How We Built It

The app integrates Snowflake's Arctic model through the Replicate API, with Streamlit serving as the user interface.

Challenges We Ran Into

One significant challenge was handling long prompts, as the model struggles with them. For effective essay feedback, we had to break down the prompts into smaller, independent segments. Additionally, there were instances where the model did not follow instructions to avoid providing direct answers, thus impacting the goal of encouraging student critical thinking.

Accomplishments That We're Proud Of

We are proud of successfully developing an app capable of maintaining coherent, on-topic conversations.

What We Learned

Our experience highlighted that for smaller LLMs, combining classical programming logic with the intelligence of LLMs often yields better results. Using smaller prompts and gated stages, classical logic can validate the LLM's output and perform necessary custom cleaning, similar to tool use in larger models.

What's Next for Learn Mentor

Learn Mentor is still a work in progress. Future improvements include making the app more robust in handling interactions, implementing user data persistence, and ensuring adherence to formatting rules.

For tutoring subjects like algebra (not included in the current version), incorporating a structured curriculum and algebraic rules is essential. Additionally, integrating tools like LangGraph to create a principled state-accessible graph would greatly enhance the app's capabilities.

Built With

• python
• snowflake
• streamlit