Canvas
Code editor
Our team

talkthrough.it 💬

Inspiration and Functionality

The best way to learn is to teach! With talkthrough.it, you can explain concepts out loud to an AI that listens and understands what you say. It analyzes your explanation, and asks targeted questions to help you strengthen your understanding. It uses user-uploaded reference material as its source, so you can effectively study for specific courses.

Technology

Frontend & Backend: Built entirely with Streamlit, handling both the UI and server logic.
AI Core: Combined Gemini with LangChain to generate smart, context-aware questions based on user-provided ground truth.
Voice Interaction: Integrated ElevenLabs for speech-to-text and text-to-speech, making the experience fully conversational.

Challenges

Tuning the system prompt to get the right AI “personality” — slightly naive but persistent enough to catch user mistakes.
Working without popular frontend tools like React made custom components (like the whiteboard and code editor) trickier to build.
And, of course... CSS (enough said).

What Went Well

The app runs smoothly and produces high-quality AI questions.
It successfully detects when you make a mistake, deviate from the reference material, or explain something inconsistently.
Integrating ElevenLabs went really well — voices are natural and human-sounding, and speech recognition is close to perfect.

What We Learned

Discovered Streamlit as a simple alternative to traditional JavaScript frameworks: it just handled the entire thing for us.
Deepened our AI development skills, especially with LangChain and multimodal interaction.

Next Steps

We had a lot more ideas we could integrate into our app, but we only had a 48 to implement!

Collaborative workspaces: create shared environments, and work together to teach a subject to an adversarial AI.
More formats for user to upload (lecture recordings, code, websites, etc.)
More flexible UI: add multiple whiteboards and code tabs