AI study APP

A freeform tablet workspace where floating study tools (chat, ink canvas, PDF reader, flashcards, quiz) share live context with an AI layer that adapts based on what the student is trying to do.

Comment

What's next for AI study APP

Here's a full draft. I've leaned into authenticity because hackathon judges have seen a thousand "we wanted to improve education" intros — the financial honesty will actually stand out.

The Problem Nobody Talks About Honestly

Studying is broken, and everyone who's been through it knows it.

You open Notion to take notes, switch to GoodNotes to annotate a PDF, jump to Anki to review flashcards, then open a calculator app, then scroll back to find where you were in your notes. By the time you've context-switched four times, thirty minutes have passed and you've retained almost nothing. The workflow isn't studying — it's managing tools.

We lived this. Long study sessions that felt productive but weren't. Flashcard decks we built and never reviewed consistently. Notes that existed but couldn't talk to anything else. AI tools that answered questions but had no idea what we were actually working on. And underneath all of it, a frustrating awareness that our ability to succeed academically — and financially — was being throttled not by intelligence, but by the friction in the tools we were forced to use.

We built AIStudyApp because we needed it ourselves, and because we believed the study tools industry had fundamentally misunderstood what students actually need.

What We Built

AIStudyApp is a tablet-first freeform study workspace where everything lives on one canvas. Think of it as a desktop operating system, but for studying — floating panels you can drag, resize, and arrange freely: a chat panel, an ink canvas, a PDF viewer, a flashcard deck, a quiz, a symbolic math calculator, and an AI layer that reads context from all of them simultaneously.

The key insight is shared context. When you highlight a paragraph in your PDF, draw a diagram on the ink canvas, and then ask the AI a question, the AI already knows about the highlight and the diagram. You don't have to paste anything, explain anything, or switch apps. The workspace is the context.

The AI adapts to you personally — your age, education level, learning style, and whether you have learning differences like dyslexia or ADHD. A high schooler and a graduate student asking the same question get meaningfully different answers. A student with ADHD gets responses structured to reduce cognitive load. The system accommodates rather than ignores.

Flashcards are generated automatically from your highlights, notes, and chat history, then scheduled using the SM-2 spaced repetition algorithm — the same algorithm behind Anki, but embedded seamlessly into your workflow instead of sitting in a separate app. The quiz module tracks topics you get wrong and builds a weak area profile that feeds back into AI responses and flashcard prioritisation. The symbolic math calculator (backed by SymPy on a FastAPI microservice) produces step-by-step LaTeX solutions similar to Photomath, rendered inline in the panel.

How We Built It

The stack is React Native with Expo's bare workflow, targeting Android tablets first. We chose this deliberately — bare workflow gave us access to native modules without sacrificing the cross-platform foundation we'll need later. The canvas is rendered with React Native Skia. Handwriting recognition runs on-device via Google ML Kit. PDF viewing uses PDF.js in a WebView (we deliberately avoided every React Native PDF library — they're all problematic in different ways).

State is split between Zustand for reactive UI state, MMKV for fast synchronous storage, and SQLite via op-sqlite for structured data like chat history, flashcards, and the RAG index. All retrieval is BM25 in pure TypeScript — no vector databases, no embedding API calls, no internet dependency for search.

The AI routing layer on the proxy server is one of the parts we're most proud of. Requests are classified by task type and complexity, then routed to the appropriate model — from free lightweight models for simple queries to Claude for complex analysis — with automatic escalation on failure, soft degradation as credits deplete, and abuse detection. The whole system runs as an 11-step pipeline on a Fastify proxy, keeping API keys off the device entirely.

On-device inference via llama.rn means free-tier users get real AI capability using their device's own compute, with no ongoing cost to us or them.

The Challenges

The honest answer is: almost everything was a challenge.

React Native's New Architecture (Bridgeless mode, TurboModules, Hermes) is powerful but the ecosystem hasn't fully caught up. Native modules that worked fine in the old architecture break silently in the new one. We spent significant time resolving a class of runtime crashes caused by MMKV being instantiated at module scope — the fix required auditing every file that touched storage and converting all instantiation to lazy getters.

The AI layer created circular dependency problems between the orchestration service, context packager, and feature gating module that required extracting shared utilities into a separate file to resolve cleanly.

Building the cost-aware routing system meant thinking carefully about what "fair" looks like when users have different budgets and different needs — and encoding that logic in a way that degrades gracefully rather than failing hard.

What We Learned

That the hard part of building AI-powered tools isn't the AI — it's the context. Getting the right information in front of the model at the right moment, without the user having to think about it, is an unsolved UX problem most AI apps haven't seriously addressed.

We also learned that constraints produce better architecture. Committing early to "no vector databases, no native PDF libraries, no images sent to the AI unless necessary" forced us to find solutions that are faster, cheaper, and more privacy-respecting than the obvious path.

Why It Matters

The tools available to students who can afford iPads with Apple Pencils and premium Anki subscriptions are genuinely better than what everyone else has access to. We want to close that gap — a capable, AI-integrated study workspace that runs well on a mid-range Android tablet, keeps costs low through on-device inference and smart routing, and actually adapts to how each student learns.

Better study tools produce better outcomes. Better outcomes produce more options. That chain matters.

Built With

Share this project:

Updates