Cognito (HackMIT 2026 Runner-Up)
AI tutor that builds a study plan from creators you trust and coaches you through homework on a whiteboard without giving away the answer
- Timeline
- Fall 2026
- Stack
- Next.js, TypeScript, tRPC, PostgreSQL, Claude, OpenAI API
- Links
- Repo
Inspiration
Khan Academy, The Organic Chemistry Tutor, and 3Blue1Brown are the creators many of us look to when we actually want to learn. But when we get stuck on homework, we open ChatGPT, paste in the question, get a wall of text back, and learn nothing. We asked: how can AI help you actually learn instead of just doing your homework for you?
Cognito is our answer, built with a team of four at HackMIT. It's an AI tutor that builds a personalized study plan from creators you already trust, then coaches you through homework on a whiteboard, reading your handwriting and giving hints without giving away the solution.
What It Does
- Personalized study plan. Tell Cognito what you want to learn and it builds a plan around it, shown as a topic graph so you can see how ideas connect.
- Lessons built from creators you trust. Each lesson pulls in existing YouTube videos and tutorials from your favorite creators.
- Video lessons with your own AI tutor. Watch alongside a tutor that knows where you are in the lesson.
- Homework mode. Upload an assignment and work through each problem on a whiteboard with an AI coach that reads your handwriting, checks your steps, and nudges you toward the answer without handing it over.
How We Built It
The frontend is Next.js, React, TypeScript, Tailwind CSS, shadcn/ui, Radix UI, and Zustand, with tldraw for the whiteboard, React Flow for the topic graph, Monaco for the code editor, and Streamdown for markdown. The backend uses Next.js route handlers and tRPC over PostgreSQL with Drizzle ORM, with better-auth handling anonymous sessions plus ChatGPT and Google OAuth.
On the AI side, Claude (via the Anthropic SDK with tool-use generation) and OpenAI GPT handle tutoring and generation, Deepgram powers the voice agent, and Desmos and Mathpix cover math. Firecrawl finds and pulls in learning resources. Jev/TypeSafe acts as a decision layer, picking each lesson's YouTube video with a typed decision instead of a generative call.
Challenges
- Quality of instruction. Generated lesson content can be shallow or wrong. We chose not to generate lessons at all and instead curate the best existing content, using AI to organize, sequence, and personalize it.
- Teaching without leaking. A tutor that just answers defeats the purpose. We had to make the AI follow traditional tutoring practice: notice where you're stuck, give a hint, and hold back the solution.
- Combining inputs into one tutor. Conversation, whiteboard strokes, and lesson progress all had to feed a single, coherent tutor with low latency.
Accomplishments
- A low-latency AI tutor that reads and understands your handwriting on the whiteboard
- Handwritten math and chemistry notation checked as you write
- Lesson plans represented as graphs, giving each custom lesson the context of what comes before and after it
- A tutor that behaves like a real one: it nudges you in the right direction and never gives away the answer
What We Learned
We learned how to build one cohesive multimodal AI, where conversation, whiteboard strokes, and lesson progress all feed the same tutor. The hardest part wasn't building the model. It was deciding what the tutor should not say.
My Contributions
I built the onboarding flow and the lesson plan graph generation and customization, which turns a learner's goal into the topic graph that the rest of the app is organized around. My teammates covered the whiteboard and homework coach (Angela), the web crawler, database, UI, and tutor voice (Evan), and lesson generation, API calls, whiteboard features, and file uploads (Joel).
What's Next
- Subject-specific tools, like a code editor for programming and CAD for engineering
- Testing how much more students learn on Cognito than with alternatives like ChatGPT
- Better resource curation and broader coverage across subjects
Cognito was also entered in sponsor challenges from Ramp, Cognition (Best Use of Devin), Deepgram, The Token Company, OpenAI, Dropbox, and Long Lake.