Pindu: Hyper-personalized graded reading with spaced repetition
Advances in generative AI have massive consequences for language learning. Pindu harnesses GenAI to bridge the gap between vocabulary flashcard practice and natural reading fluency by building bespoke reading passages and integrating reading with SRS. It's an evolution for me as a long-time Anki user and language learner.
Finally, a usable implementation of my ideas
Working toy showing the project's ideas
The "Sentence Suggester" Project
Pure spaced repetition vocabulary review for language acquisition (using, e.g., Anki) is boring! And even if you have the discipline to sit through hundreds of card reviews each day, you're leaving learning outcomes on the table by not reviewing the vocabulary within a natural language context. Yet, a scheduled and tracked review system gives back the motivation of visualized progress towards goals and the comfort of knowing you're "on track" and not wasting time. Other methods of language acquisition, such as graded readers or pre-recorded and subtitled audio/video, give the most natural learning experience, but they also frustrate with their non-optimal and impersonal approach to content. What if these approaches could be combined? They can. (Obviously.)
I began pondering that question in 2020 after years spent studying thousands of Anki flashcards for Chinese, totalling hundreds of hours (no joke!) doing recall word-by-word. At the time, just on the cusp of GPT-3's debut, I hatched a plan to brute force through scraped Chinese web content to surface natural Chinese sentences that contained a high density of due-to-be-reviewed flashcards: a review of a single sentence would count as a review of all of its constituent words. I started tinkering with a possible Anki extension to allow this. By the end of 2020, I was working on a standalone R Shiny application (the only web framework I knew) with the OpenAI API to create novel sentences on the fly. [Life happens.] At the end of 2024 I picked up the project again. Not only had LLMs exploded in capabilities, but I could now plausibly write software applications for myself, aided by coding assistants like Aider and Claude. And now at the beginning of 2025, my dream seems within reach...
Linked posts:
Major revisions:
- 2024-12-31: First imprint, to publish my "demo app"
- 2025-03-18: Reworking the project along with site maintenance