Check out this project here and the tutoring side here
Building my dream edtech platform
Ever since I started high school I have been trying to improve the way I learn, both in how efficient it is and in how well I actually understand things at the end of it. Over the years I tried spaced repetition, mnemonics and a lot more, and by the end of high school I had a pretty good system in place. By the time I went to university AI models had finally got to the point where they were actually useful for learning, so I wanted to bring them into that system. The first thing I did was simply get the AI to make flashcards from my lecture notes. This worked well and made sure I had consolidated and understood what was taught in each lecture but I knew there was more I could be doing.
The main inspiration for this project came from an episode of the Dwarkesh Podcast with Andrej Karpathy. In it Andrej talks about learning Korean and how good his Korean tutor was, and I think he managed to describe exactly what a perfect tutor is. A perfect tutor serves a student exactly the right level of question. The right level of question can't be too hard that the student just gets stuck on it. It also can't be too easy that it's trivial. Using this principle a tutor builds what he calls "ramps" to knowledge, where each part of the ramp gets harder as the student's ability grows. Doing this well means the tutor has to keep an accurate picture of what the student knows, which is the part I thought software could do.
What cookie does
Cookie is an all in one revision app that learns your level. It's for students doing any exam around the globe or just studying for their own pleasure. The fundamental idea behind cookie is that the AI learns your level of understanding so can serve you material to do based on that. You can give it your own notes or any other materials and it writes full exam-style questions from that material. It can then also mark the answers for you or you can still mark it yourself. If you think the AI marked it wrong you can give it suggestions to correct its mistakes. There are also flashcards which I made sure can be imported from and exported to Anki to make transition into cookie as easy as possible. There are also past papers, notes and a revision plan built around your exam dates and an assistant you can ask about any of it.
How the product works
Everything in cookie is based off topics these topics come from the content and use a subtopic structure for example maths/calculus/differentiation. Any content anyone uploads gets its own topics made automatically. Whenever someone uploads new content we first try and match it to a topic that already exists. Then if nothing matches the AI is used to pick one. This means the same topic does not end up split across a bunch of slightly different names.
When anyone does an action on the app i.e. answers a question, rates a flashcard or marks a past paper question the app makes a judgment about that topic and it is added to their record. I split these into hard evidence such as marked answers and card ratings but also soft evidence like the assistant noticing a misunderstanding or a student asking for a hint on a question. The soft evidence is capped so that it can never decide on its own that a student knows or doesn't know something. Whenever the record is read it works out a score for each topic rather than permanently storing one. The older the evidence is the less it counts for with a half-life of about 90 days. This means that topics you have not touched for months are shown to you instead of just assuming they are still fine. For topics that have never been tested instead of having a score of zero they are marked as being "never tried". Evidence for a particular subtopic also counts towards its parent but less significantly.
Every part of the app then reads from this one record. This includes: which question you get next, what is due today, the revision plan, your predicted grades and the AI assistant. For the flashcards we also use the FSRS spaced repetition algorithm, and once someone has done enough reviews cookie fits the algorithm's parameters to their own memory every night. Due to the fact a lot of users may be under 18, the softer guesses the AI makes about a user are switched off by default for users who are under 18. They can also see everything cookie has recorded about them and tell it when it is wrong.
The tutoring side
When I started university I also began tutoring on the side to earn a little extra money. After I started building cookie I realised that a lot of what I was building could really help my students and could be used inside my lessons. So I decided to add a tutoring marketplace, Cookie Tutoring, which gives tutors and their students free access to cookie. This was also a good way to help monetisation, as students famously don't have much money, and in my experience only the most dedicated ones will pay for a subscription out of their own pockets however good the product is.
As well as the marketplace I built a full video platform to host the lessons, with live video, a shared whiteboard and payments on the platform. A tutor can set work in cookie and follow a student's progress between lessons, and I have added a few other AI features to help tutors tutor better, like an AI lesson planner and AI suggestions for lesson reports. Because tutoring often involves children a lot of the work on this project was actually legal stuff instead of code. This meant another benefit of this project was that it allowed me to learn a lot about the law throughout it.
How it is built
Cookie and Cookie Tutoring are two separate apps and are both written in Next.js and TypeScript and are hosted on Vercel. They share one Supabase (Postgres) database and also one login. This means one account works on both. For the video for the lessons we use LiveKit and for the whiteboard we use Excalidraw. For payment we use Stripe.
All of the AI goes through a custom harness of its own rather than each app calling the model providers directly and every AI action costs credits. I built almost all of this by directing AI coding agents like Claude Code and Codex. This was the only way a project this size was possible for one person. Particularly as I'm studying my degree alongside it. I also think this is just the way software is going now so doing this I think was really valuable as it taught me how to direct and use these agents effectively. As I said in my construction site post, I think this only works because I learnt the fundamentals first. During this project I mostly spread my time across three things: designing the architecture of the product, checking code written by the agents and testing the code in a browser.
What's next?
I have lots of ideas for where to take cookie next. One of the main challenges in learning, which unfortunately cookie does not solve yet, is motivation. Even with good software you still need the motivation to go and actually use it. I have a few ideas on how to help with this, but a product I really like the look of is Habits First by Reysu, so either integrating with it or building something similar into the app is a priority at the moment.
Another thing I would like to add is a much cheaper subscription where you can use your existing AI subscriptions or API keys instead of cookie's own, so people are not paying for a bunch of different subscriptions. I would also like to connect cookie to personal agents like Meta's Muse or GrokBot, so that you could use that subscription for cookie as well, and so those agents could read your knowledge record and make learning decisions for you.