An emotionally adaptive offline AI learning companion. One Gemma 4 E2B model running locally via LiteRT-LM — stories that read like a WhatsApp group chat, characters pulled from 80 of your favourite franchises, and a Learner Twin that actually remembers you.

🚫 No cloud calls 🚫 No API keys 🚫 No data leaves the device ✅ 100% private, fast, offline-first

Built for the Kaggle Gemma 4 Good Hackathon 🌍✨ (Future of Education)

Most AI education apps depend on the internet. That means they fail exactly where learning matters most:

📶 Rural schools with weak signals 🚫 Campuses with blocked Wi-Fi 📱 Students surviving on 2G data 🚌 Long commutes on buses or trains 🏝 Remote villages and islands 🔌 Places where connection drops daily

Once installed, it works forever offline. Your tutor, planner, storyteller, and study companion live inside your phone ❤️📱

The flagship moment. Type any topic, pick a vibe, and the lesson arrives as a real chat conversation between 2 characters reacting, joking, paraphrasing, getting it wrong, then getting it right.

🎭 Two modes:

• 🛠 Practical — friend-group chat, daily-life examples ("bro why do bikes skid in the rain?" / "less friction" / "wait so friction is actually GOOD??")
• 🎬 Movie / TV — pick from 80 franchises (anime, cartoons, K-drama, gaming, Indian TV, movies). Naruto, Stranger Things, Money Heist, Goku, Walter White, MS Dhoni, Itachi… they teach in their own voice.

⚡ Wave-based generation: the first 6 scenes stream in, then up to 2 more continuation waves push the lesson to 16–18 messages — depth without the small model getting overwhelmed by a single mega-prompt. Each wave re-injects the persona block so cast voices don't drift.

🔁 Each message auto-reveals on a 1.7 s typing-pill cadence (like a real chat), scrolls into a chat feed, and the cast + speaker order is locked in code so Gemma can't drift off-character or monologue as one person. We also filter out mascot characters (Pikachu, etc.) at cast-build time — they have rich voice samples but can't carry a teaching dialogue in English.

Your study companion that builds a private model of YOU over time, fully on-device.

Ask:

📝 What should I study next? 😵 Where am I weakest? 📈 How am I improving? 🔥 Motivate me today

Every chat compounds — it doesn't reset between sessions. Every quiz, every mood check-in, every story you finished feeds the same SQLite memory.

Type any topic into Learn Anything and the default action now builds a Duolingo-style stepped path: 4 steps at basics, 7 at intermediate, 11 at advanced. Each step has concept tags + a difficulty badge, and ticks fill in as you pass at ≥70% accuracy.

Each step is one short stepped story. Want a single throwaway chat instead? The chip below the input — "Skip the path · just chat about it →" — bypasses the path and goes straight to a one-shot lesson.

🪶 Mastery generation has belt-and-braces JSON parsing: tight prompt → schema-shape validation → retry with key-names spelled out → last-resort prose salvage that extracts "Stage N: title" headings from any markdown the model emits. The user never sees a dead end.

After a strong score (≥70%) on a franchise lesson, the tables flip — the character admits they didn't quite get it and asks you to teach them. Three deterministic turns (opening question → follow-up → lightbulb moment), all in their authentic voice. Teaching is mastery.

Daily mood check-in (calm / hyped / curious / anxious / sad) quietly reshapes the tone of every story and the picker's franchise suggestions — anxious days lean toward calmer, gentler casts; hyped days unlock loud bombastic ones.

• Dyslexia-friendly mode — Atkinson Hyperlegible font, BionicText word emphasis, prompts auto-shorten sentences.
• Read-aloud (TTS) — every chat bubble can be spoken aloud with karaoke-style word highlighting via the device's offline TTS engine.

Both default OFF. Zero footprint when not used.

No tracking. No ads. No cloud syncing. No surveillance.

Your learning belongs to you ❤️

# Deps
flutter pub get

# Code generation (Riverpod / Freezed / JSON)
dart run build_runner build --delete-conflicting-outputs

# Run on a connected device
flutter devices
flutter run -d <DEVICE_ID>

# Optional: pass a Hugging Face token for in-app download
flutter run -d <DEVICE_ID> --dart-define=HF_TOKEN=<token>

1. In-app download (recommended) — tap Download Gemma 4 E2B on first launch. The download runs in the background via flutter_downloader with a system notification, survives screen-lock / app-backgrounding, and resumes on partial network drops. On completion the engine auto-installs and warms — no second tap.
2. Sideload via ADB — if you already have the .litertlm:

   adb push gemma-4-E2B-it.litertlm \
     /storage/emulated/0/Android/data/com.vidyasetu.vidyasetu/files/

   First launch detects the file and silently imports + warms it. You'll see the new BootstrapScreen (branded splash + progress bar + status text) rather than the old mute black splash.
3. Pick from device — if the file is sitting in Downloads or shared from another app, tap "Pick a .litertlm file from your device" on the setup screen. Works for files anywhere image_picker / file_picker can read.

✅ Android 12+ (minSdk = 31, hard requirement for LiteRT-LM) ✅ ≥ 4 GB RAM ✅ ≥ 5 GB free storage (model is ~2.58 GB + temp copy during install)

1️⃣ Pick a model: background download (~2.58 GB), ADB sideload, or browser-download + file picker 2️⃣ Wait through the BootstrapScreen (progress bar + status) while LiteRT-LM warms the engine — ~25 s on a mid-range Android, ~50 s if it's also copying from sdcard 3️⃣ Enter name + grade + language 4️⃣ Optionally toggle dyslexia-friendly mode + read-aloud 5️⃣ Done forever ✅

After that:

✈ Turn on airplane mode 📴 Disconnect Wi-Fi 🌍 Go anywhere

Learnify still works perfectly.

1. Mood check → tap Curious
2. Learn Anything → type "why the sky is blue"
3. Movie / TV → pick a franchise that matches your mood
4. Watch the chat unfold scene by scene
5. Ace the quiz → unlock TEACH ${CHARACTER} and flip the table

lib/
├── core/
│   ├── ai/           # GemmaService, GemmaOrchestrator, AgentPrompts
│   ├── db/           # SQLite (profiles, topics, quiz, memory, paths)
│   ├── franchises/   # FranchiseLoader + persona models (80 franchises)
│   ├── services/     # LocalProfileService, LocalMemoryService, TTS
│   ├── theme/        # Dark glassmorphism + dyslexic font variant
│   └── widgets/      # GlassContainer, NeonButton, BionicText, KaraokeText
├── features/
│   ├── setup/        # Model download / sideload + profile creation
│   ├── auth/         # Home shell + dashboard
│   ├── story_learning/  # Chat-bubble Story + Feynman screen
│   ├── companion/    # Learner Twin chat
│   ├── mastery_path/ # Duolingo-style stepped UI
│   ├── scan/         # Multimodal textbook capture
│   └── profile/
└── routes/           # /lesson, /feynman, /scan, /mastery-path, /home/*

One Gemma model in RAM. Identity = system prompt. The orchestrator routes intent and threads language + mood + dyslexic flags into every call.

User → GemmaOrchestrator
   ├── 📖 Story            (16–18-scene chat lessons via wave continuations)
   ├── 🧠 Tutor            (concept explanation)
   ├── ❓ Quiz             (5 questions w/ concept tags, weak-area aware)
   ├── 📅 Planner          (7-day study schedule)
   ├── 🔍 Explorer         (sub-topic decomposition)
   ├── 🗺 Mastery          (4 / 7 / 11 step paths, JSON-strict + prose salvage)
   ├── 👤 Learner Twin     (Companion chat, names recent weak concepts)
   ├── 📷 Scan Textbook    (ML Kit OCR → text-only Gemma; multimodal isn't viable on the .litertlm E2B build)
   └── 🎓 Feynman          (role-reversal teaching)

⚡ Shared weights — one model, all agents ⚡ Progressive streaming on Story (intro paints in seconds, tail arrives in background) ⚡ Cast locked in Dart so Gemma can't invent off-character names ⚡ Add a new agent = a new prompt + an orchestrator method

Education should not depend on internet speed.

A child in a village deserves the same AI tutor as a student in a smart city. A learner on a train deserves help even without signal. A classroom deserves intelligence without surveillance.

Learnify brings equal opportunity through offline AI. ❤️

Kaggle Gemma 4 Good Hackathon Theme: Future of Education

Built on Gemma, subject to Gemma Terms of Use.

The future of education should fit in your pocket — and feel like a friend. 📱❤️✨