ThinkStack

Inspiration

I've never been a fan of learning with a textbook or by listening to long, boring lectures. However, hands on learning experiences are difficult to find with many subjects because building them takes time and expertise most teachers don’t have. A simulation, a graph students can manipulate, or a coding exercise that actually runs has to be designed, built, and tested for every single lesson. So most classrooms fall back on what’s easy to produce: slides, worksheets, and textbook readings. The students who would learn best by trying things end up watching and memorizing instead.

We built Thinkstack to close that gap. An instructor describes what they want to teach, and AI drafts the whole lesson with the interactive pieces already built in. Students read a short explanation, then immediately play with a simulation or run their own code, and a tutor is there to guide them without handing over the answer. Hands-on learning shouldn’t depend on how many hours a teacher has to build it.

What it does

ThinkStack allows users to create, design, and publish their own lesson plans and courses easily with the help of AI. ThinkStack AI incorporates an interactive, hands on learning process using simulations, graphing, quizzes, flashcards, etc. You name it, we'll create it with AI. Every lesson enables readers to ask a personlized AI tutor for help at each step. It can access the lesson, see what your stuck on, and give you hints without revealing a solution. This software is meant to empower educational institutions around the world to improve the quality of learning with AI. ThinkStack makes it easy for Instructors to create interactive, hands on activities for students, create a course, and add students to their course.

How we built it

We used MongoDB to store all data related to lessons, courses, and user information for authentication and account creation. We used a combination of Gemini and Backboard API for our LLMs with a reliable fallback system that incorporated over 10 different models, so if any one failed the app would still run smoothly. We used Godaddy and Vercel with Nextjs, React, and typescript for the tech stack of the web application which included backend software, UI, and deployment.

Challenges we ran into

Challenges included secure data storage, authentication, and efficient model routing and workflows to generate high quality outputs. The software incorporated RAG and model memory along with system prompt design and agent hierarchies.

Accomplishments that we're proud of

We are proud of a lesson you can actually teach from. A teacher types a topic, a level, and how long they have, and Thinkstack writes a full page: sections, inline equations, a check at the end of each section, and a simulation or runnable example when the idea needs one. A wrong answer turns red and the explanation stays hidden until the student gets it right. Anyone can open a public lesson with no account and ask a tutor that marks the passage on the page and will not pick the quiz choice. Instructors can publish a course, approve who joins, and keep the roster.

What we learned

Writing the lesson was the easy part. Making it survive contact with a model was the work. Serverless requests die at 60 seconds, so generation has to plan first and then fill one section or one interaction at a time. Models eat LaTeX, swallow markdown, and draw simulations upside down unless the prompt is strict about the frame and the coordinate system. A tutor will also paraphrase a quiz answer unless you tell it, in plain words, not to pick the choice.

What's next for ThinkStack

A class view of how students did on the checks, then due dates on course lessons. Faster generation, so a teacher can watch a lesson appear instead of waiting on a finished one. Tighter interactions, so every simulation fits the page and matches the idea it is there to show.

Built With

• backboard
• cursor
• geminiapi
• mongodb
• typescript