🦦 Study Quest

Learn Reinforcement Learning, LLMs, System Design and AI Safety by doing, not just reading.

Study Quest is a desktop app that turns five long-form books and two hands-on guides into lessons you can read, run, sketch and quiz yourself on. It has live marimo notebooks, a coding playground with tested exercises, Excalidraw boards for system design, and Byte the otter 🦦, who nudges (and roasts) you until you open a lesson.

Built with Tauri 2, React and Rust. Runs on macOS, Windows and Linux. Everything stays on your computer.

---

✨ What's inside

Four learning paths, one at a time

Path	Books	Topics
🎯 RL	Reinforcement Learning · RL Environments & Verifiers	201
🧠 LLMs	LLM Fundamentals · Pre-Training to Post-Training	145
🏗️ System Design	System Design: Zero to Expert · DSA and ML Interviews	54
🛡️ AI Safety	Alignment · Interpretability · Scalable oversight · Truthfulness · Adversarial robustness · Safety evaluation	30

Pick a path in the sidebar and only that path is shown, with a progress ring, chapter list and a Continue button. Lessons keep their math (KaTeX), diagrams (compiled from TikZ) and code.

RL Environments & Verifiers is the practical side of RL for LLMs: building environments, designing verifiers that resist reward hacking, sandboxes for agentic rollouts, and training-data pipelines that survive a real run.

Learn by doing

• ▶ Runnable lessons: every Python block has Run / Edit, notebook-style: earlier cells run first, output and plots appear inline.
• 📓 Notebooks in the page: reactive marimo notebooks open inside lessons. 23 interactive ones (bandits, PPO clipping, attention heatmaps, KV-cache calculator, reward hacking, calibration, linear probes, judge bias, prompt injection, verifier and sandbox labs…) plus every book script. Create your own; they live in ~/.study-quest/notebooks.
• 🧪 Coding playground: 60 tested exercises across DSA, ML math, transformers, RL, inference, system design, agents, RL environments and AI safety, plus a free scratchpad.
• 🎬 3D explainer videos: worked-example animations for the major topics: bandits, MDPs and returns, value iteration, TD vs Monte Carlo, policy gradients, PPO, RLHF, GRPO, embeddings, attention, the transformer block, next-token prediction and training on a Videos page and inside the lessons they explain.
• ✏️ Design it yourself: every System Design and DSA lesson has an Excalidraw board, saved per lesson. All your boards are on the Diagrams page; export as .excalidraw.

Make it stick

• 🔮 Warm-up: rate what you know and guess before reading.
• 🧠 Quiz me: derivations reveal one step at a time; key ideas hide until you've recalled them.
• ✍️ Explain it back: write it in your own words; the app checks which key terms you used and compares with your guess.
• 🔗 Connect it, 🌫️ What's unclear? (shown on your dashboard until resolved) and 📈 confidence before vs after.
• 🎙 Voice typing on every box (falls back to 🌐 twice on macOS, Win + H on Windows).

Keep coming back

• Streaks, XP, levels, badges and an activity card (this week vs last, daily-goal chart, heatmap).
• Desktop nudges with a troll level from 😇 Nice to 🔥 Unhinged. At the top end Byte brings up your crush's LinkedIn and your screen-time report.
• Plan my week: books study blocks on weekday evenings. macOS adds them to Calendar (✕ removes them again); Windows and Linux open an .ics in your calendar app.
• 🎨 Themes: System / Light / Dark and eight accent colours (or pick your own).

---

📦 Install

Download from Releases: a versioned release, or nightly for the latest commit:

OS	File	First launch (builds aren't signed yet)
macOS (Apple Silicon / Intel)	.dmg	See macOS first launch below
Windows 10 / 11	.exe installer or .msi	SmartScreen: More info → Run anyway
Linux x86_64 / ARM64	.AppImage or .deb	chmod +x Study*.AppImage && ./Study*.AppImage, or sudo apt install ./study-quest*.deb
Linux ARMv7 (e.g. Raspberry Pi 32-bit)	.deb	sudo apt install ./study-quest*armhf.deb

Python for exercises and notebooks. In Settings → Python, pick one:

• Set up with uv (recommended): one click creates the app's own environment in ~/.study-quest/.venv with the pinned packages from requirements.txt (marimo, numpy, matplotlib; optionally PyTorch). Needs uv: brew install uv, winget install astral-sh.uv, or curl -LsSf https://astral.sh/uv/install.sh | sh.
• Use my own Python: paste any interpreter path (a conda env, a project .venv, pyenv…) and press Test. It shows the version and which packages it has. Install what's missing with python -m pip install -r requirements.txt.

The app uses, in order: your custom path, then the uv environment, then the system Python.

Platform differences

	macOS	Windows	Linux
Calendar	Synced to a "Study Quest" calendar	.ics export	.ics export
Sidebar	Translucent (vibrancy)	Solid	Solid
Tray / menu bar	Menu-bar streak 🔥	System tray	Needs an AppIndicator desktop; otherwise closing the window quits
Dictation fallback	🌐 twice	Win + H	Desktop dictation tool

Linux needs WebKitGTK 4.1 (preinstalled on Ubuntu 22.04+, Fedora 37+).

---

🛠 Develop

Requirements: Node 24+, pnpm 10, Rust (stable), uv (it fetches Python for you). On Linux also the Tauri system libraries: sudo apt install libwebkit2gtk-4.1-dev libayatana-appindicator3-dev librsvg2-dev.

pnpm install
uv sync                                       # ./.venv with the pinned Python packages (add --extra torch for PyTorch)
pnpm tauri icon src-tauri/icons/source.png   # once: generates the app icons
pnpm tauri dev                                # run with hot reload
pnpm build                                    # type-check + build the frontend
pnpm tauri build                              # installers for your OS

The book content (public/content/, public/figs/) and notebooks/ are generated from LaTeX sources kept outside this repo, and committed here so anyone can build the app. The AI Safety and RL Environments guides are Markdown in guides/, built by scripts/build_guides.py.

Path	What
src/views/, src/components/	React pages and UI
src/lib/store.ts	All saved progress and settings
src/lib/platform.ts, src-tauri/src/platform.rs	Everything that differs between macOS, Windows and Linux
src-tauri/src/	Rust: notebooks folder, marimo server, Python runner, Calendar, nudges, tray
scripts/exercises_src.py	All playground exercises (source of truth)
scripts/build_notebooks.py	Example marimo notebooks
guides/*.md, scripts/build_guides.py	Markdown guides (AI Safety, RL Environments)
videos/web/scenes/*.js, videos/record.mjs	3D explainer scenes and the recorder that renders them to public/videos/
src/nudges.json	Every notification and roast
AGENTS.md	Conventions for AI coding agents (and humans)

Every push to master refreshes the nightly pre-release. Pushing a tag like v0.2.0 publishes a versioned release with every installer.

macOS first launch

The builds are ad-hoc signed, not notarised by Apple (that needs a paid Apple Developer account), so macOS blocks the first launch. Drag the app to Applications, then either:

• open System Settings → Privacy & Security and click Open Anyway next to Study Quest, or
• run xattr -dr com.apple.quarantine "/Applications/Study Quest.app" once in Terminal.

Building the .dmg yourself: pnpm tauri build skips create-dmg's Finder step (it only arranges the icons in the DMG window and needs Automation permission, otherwise it fails with Not authorized to send Apple events to Finder). Set STUDY_QUEST_DMG_LAYOUT=1 to keep it.

---

💜 Contributing

The most useful thing you can add is a topic. If you know something a learner on the RL, LLM, System Design or AI Safety path should know, please share it. PRs of any size are welcome, and so are issues that just say "this topic is missing".

Suggest or write a topic

Open an issue with the topic label, or a PR. For AI Safety and RL Environments, add a ## Topic straight to guides/ai_safety.md or guides/rl_envs.md and run uv run python scripts/build_guides.py --check (it runs every Python block). Other paths: a Markdown or LaTeX file under contrib/topics/. A good topic:

1. covers one idea, with a real-life analogy;
2. builds any math up step by step (each step becomes a reveal in Quiz me);
3. puts key terms in bold (Explain it back checks for them);
4. ends with a small runnable Python example and 2–3 key ideas.

Topics we'd love: newer RL-for-LLM methods, inference systems, evaluation, safety and alignment, distributed training, system design case studies, and clearer explanations of anything that confused you.

Add a coding exercise

Add an E(...) entry to scripts/exercises_src.py (prompt, starter, solution, test_* tests, hints, topic= keyword), then run:

uv run python scripts/build_exercises.py

It refuses to write unless the solution passes every test and the starter fails at least one.

Add a notebook

Add an entry to INTERACTIVE in scripts/build_notebooks.py (sliders and plots welcome), give it a group in GROUPS and a topic keyword, and check it runs with marimo export html.

Add a 3D explainer video

Copy a file in videos/web/scenes/, export meta (title, path, duration, keywords that match lesson titles) and setup(stage) returning renderAt(t), and add its id to SCENES in videos/web/main.js. Preview with cd videos && npm install && npm run serve, then render with npm run record <id> (needs ffmpeg and Google Chrome). Keep the light background and keep each video around 30 s.

Add nudges and roasts

Add lines to src/nudges.json. Keep them funny and about studying: tease habits (scrolling, "I'll start Monday", 47 open tabs), never appearance, identity, health or anything that would actually hurt. ML puns score double.

Pull requests

• One topic, exercise or fix per PR.
• pnpm build must pass; CI builds macOS, Windows and Linux on every PR.
• Say what a learner gains from the change.

---

🔒 Privacy

Everything stays on your computer. Progress is saved locally (export or reset it in Settings). Code runs with your own Python in a temp folder with a time limit. marimo listens only on 127.0.0.1 with a random token and stops when you close the window. On macOS, Calendar events are created through Automation, which asks your permission first.