GroundEd
Built for #BuildHybridAI 2026 · Firebase AI Logic hybrid inference

The classroom AI that works when the internet doesn't.

GroundEd is an offline-first teaching kit for low-connectivity classrooms: quizzes, lesson plans, flashcards, live sessions, paper-stack grading and class insights — all from the same material. Every AI action runs through Firebase AI Logic hybrid inference — Gemini Nano on the device when it's there, Gemini in the cloud when it helps — and answer keys are verified on the other tier, so two independent brains check every question.

Open the Studio
Student? Join your class:
ON DEVICE · private · free · airplane mode CLOUD · long documents · heavy lifting Student answers stay on the device whenever possible

Two tiers, one honest badge

On-device tier

ON DEVICE

Gemini Nano through Chrome's built-in AI, driven by InferenceMode.PREFER_ON_DEVICE. Quiz generation and answer grading keep working in airplane mode — private and free, because they never leave the device.

Cloud tier

CLOUD

The same JSON schema is wired to both tiers — the cloud Gemini model takes over for long documents and heavier reasoning, or whenever no local model exists. One schema, two wiring paths (responseSchema / responseConstraint).

The badge is driven by the SDK's own response.inferenceSource — never a guess.

From textbook page to live class — in three steps

1

Bring your material

Paste a passage or photograph a textbook page. Pick grade level 3–12, Bloom's depth and the question mix.

2

Generate anywhere

A quiz with a cross-verified answer key, a full lesson plan, and a flashcard deck — on-device when possible, cloud when it helps, badged honestly either way.

3

Run the class — live or on paper

Students join with a code on their phones, or answer on paper and the teacher types the stack in once. MCQs auto-check; short answers are graded by AI; offline answers queue and sync; results land on the Insights dashboard.

Everything a low-connectivity classroom needs

Quiz studio

Paste text, import a .txt/.md/PDF or photograph a textbook page; pick grade level (3–12), Bloom's depth and question mix — get an editable quiz with a cross-verified answer key. PDFs are extracted in the browser — nothing uploaded.

Lesson plans

The same material becomes a timed 40-minute plan: objectives, what you do / what students do per section, homework, and support & extension notes. Printable.

Printable worksheet

Real print CSS: name and date lines, MCQ option grids, short-answer rule lines and a separate answer-key page. Paper is the most reliable classroom device.

Live sessions

A 6-character join code brings the class in on their phones. MCQs auto-check deterministically (students see their score instantly); short answers are graded by AI with a per-call routing badge.

Paper mode

No student devices? Print the worksheet, then type the class's answers once — MCQs auto-check and short answers are graded by the same privacy-biased engine, offline if needed. The audience's most common medium, supported end-to-end.

Differentiated versions

One click re-builds the same material as a Support version (two grades simpler) and an Extension version (two grades deeper) — because real classrooms are mixed-ability.

Flashcards, with spaced repetition

A revision deck from the same lesson — scheduled by the SM-2 algorithm (the one behind Anki): failed cards come back tomorrow, easy cards in weeks. Progress lives on the device. Cards can be read aloud for younger readers.

Read-aloud

Questions, flashcards and re-explanations can be listened to with on-device speech synthesis — for younger and low-literacy learners. Free, offline, no AI call needed.

Class insights

Sessions, student results, MCQ trends and the questions the class keeps getting wrong — computed locally from history. Plus a hybrid-value meter: what on-device routing saved you, in calls, KB and offline completions.

Cross-tier verification

Answer keys generated on-device are re-derived in the cloud, and vice versa. Two independent brains check every key — a systematic mistake on one tier gets caught by the other.

Routing inspector

Every AI call is logged with the SDK-reported inferenceSource, model, latency and offline status — the badge is never a guess.

Offline-first sync

Answers submitted offline queue on the device (Firestore offline persistence) and sync on reconnect. Quizzes export as files — teacher-to-teacher sharing needs no server.

Built for the classrooms the cloud forgot

Only half of the world's primary schools have internet access for teaching (UNESCO UIS, 2025), and 2.2 billion people remain offline (ITU, 2025). Per-call API pricing is invisible budget poison for schools. GroundEd keeps the teacher working through all of it: the on-device tier is free and private, and the whole app ships as static files that keep running offline.

50%

of the world's primary schools have internet for teaching — UNESCO UIS, 2025

2.2B

people are still offline — ITU Facts & Figures, 2025

Under the hood — the Firebase stack

Firebase AI Logic — hybrid inferenceGemini Nano on-device (Chrome built-in AI)Gemini cloud fallbackFirebase AuthenticationApp CheckCloud Firestore — offline persistenceRemote Config — prompts + routing policyFirebase Hosting

Try it with the network off.

Open the Studio, load a sample, hit Generate — then flip on airplane mode and keep going. Works best in Chrome (the on-device tier uses Chrome's built-in Gemini Nano; the cloud tier works everywhere).

Open the Studio