web LLMs coming soon

Building a Quizlet Replacement

AI generated flashcards with coursework and notes as context

The Story

I’ve always been obsessed with speeding up repetitive processes. In college, I built a bot to “snipe” study room reservations in the undergrad library. It was a chore to do them manually because reservations opened up a week in advance, just minutes after midnight.

Mindigo started with a similar itch, and a question: why is making flashcards still a manual, soul-crushing task?

The idea actually came to me in my second year of high school. I hacked a together a command line tool that took the term and would populate a few browser tabs with information on that topic. I tried rudimentary text summarization APIs but they sucked and were too expensive (this was 2018 after all).

With the wide availability of LLMs, and especially cheap multimodal ones like the Gemini Flash models, I saw an opportunity to build what my younger self would’ve dreamed of.

I made a faster interface - one where I could drag in my raw materials (PDFs, screenshots, even handwritten notes), and get clean, editable flashcards out in seconds.

Although there is definitely educational upside to making flashcards by hand, I find myself more motivated to study when flashcards are already made. Plus, it’s incredibly cheap to do this!

The Stack

  • Frontend: Vite + React + Typescript + Tailwind

  • Serverless Functions, Auth, Postgres: Supabase + Typescript

  • LLM Inference: OpenRouter (using Gemini 2.5 Flash and GPT-4.1 mini)

Priorities

  • Speed = UX. Upload to flashcard conversion in under 20 seconds. Search by document content. Intuitive keyboard shortcuts.

  • Designed for power users. Everything’s optimized for minimal friction and flexibility. Combine photos, screenshots, and pdfs. Combine notes from different classes. Upload a whole chapter. It all works.

  • It’s gotta look good. Maybe not a top priority, but what’s a study tool without dark mode?

Discoveries

  • Gemini Flash 2.5 is a beast at transcribing anything into markdown. It worked with 80 page slide decks and my handwritten calc notes (which it converted to LaTeX). Seriously reminded me of the magic I first felt when using davinci-002 for the first time.

  • Storing transcriptions in Postgres instead of S3 might’ve been motivated by frugality at first, but it makes search and flashcard generation significantly lower latency with minimal quality loss in most subjects.

old demo video until I fix rate limits and deploy:
00:00