This is a submission for the Gemma 4 Challenge: Build with Gemma 4 I run Educloud Academy, a learning platform focused on cloud and AI skills for African students. Over the last year my team and I have done university outreach across 10 Cameroonian universities ,including my own alma mater, the University of Buea, running on-the-ground sessions on how to break into cloud and AI careers. Outreach evidence (LinkedIn): (Photos from these sessions attached below this post.) Two problems kept showing up at every single campus, no matter the topic: EduCloud (the app) is my answer to both. It's a fully offline, on-device AI study assistant that turns a learner's own PDFs, textbooks, and lecture screenshots into interactive study materials — multiple-choice quizzes, flashcards, multi-lesson workshops, mind maps, and summaries — without a single byte ever leaving the device. Because Gemma 4 is genuinely multilingual, the same binary that serves an English-speaking student in Yaoundé serves a French-speaking student in Douala without an extra translation hop. ultraLean / lean / balanced / full ) and adjusts token budgets, RAG context size, and max question count so it runs cleanly on the iPhone 13 Pro Max I tested on as well as on lower-end mid-range Androids that students actually own.Everything is built in Flutter + Dart, with native Gemma 4 inference via the flutter_gemma plugin (LiteRT-LM under the hood), and persistence via ObjectBox. iOS, Android, macOS, Linux, and Windows are all supported from one codebase — important because campus device fleets aren't uniform. Watch the walkthrough on YouTube: https://youtu.be/dVYz8xq2L_8 Full source: https://github.com/trey-rosius/Local-Educational-App Key directories to explore: lib/services/educational_tool_service.dart — all 8 function-call schemas (quiz, flashcards, workshop, plus 5 interactive tutor tools).lib/services/study_material_service.dart — generation entry points + the JSON repair pipeline + semantic validators.lib/services/rag_service.dart — PDF/image ingestion, batched embedding, HNSW search.lib/services/background_generation_service.dart — the non-blocking task queue that branches between tool-call and text-mode paths.architecture.excalidraw — full visual architecture diagram (open at https://excalidraw.com).README.md — complete feature + architecture + failure-mode documentation.Model chosen: Gemma 4 E2B (.task ). E2B was the only viable choice for a truly on-device application of this scope: E2B hits the sweet spot for what EduCloud needs: Message.withImage(...) — critical for the "snap a textbook page" feature.ToolChoice.required
- a Tool schema, structured generation is guaranteed valid — no JSON parsing failures possible. Before migrating to tool calling I had to maintain a ~700-line state-aware JSON repair pipeline to handle every quirk the model produced (asymmetric quotes, missing braces, Python-style single quotes, \X escapes outside strings, 0.0 as answerIndex instead of 0 , citation paste-throughs…). After the tool-calling migration, that pipeline is now a fallback that almost never fires.How Gemma 4 is wired into the app, end to end: flutter_gemma 's Embedder.generateEmbeddings(List<String>) batch API. Ingestion runs ~10× faster than per-chunk thanks to batching.Float32List(512) embeddings in ObjectBox.model.createChat(supportsFunctionCalls: true, tools: [...], toolChoice: ToolChoice.required) returns a FunctionCallResponse with already-parsed Map<String, dynamic> args. The runtime constrains generation at the token level.chat.generateChatResponseAsync() so the user sees Markdown materialize in real time.Message.withImage with the same Gemma 4 model — no separate OCR engine needed.chat.close() after each generation releases the KV cache; the model is reused as a singleton across tasks (closing it triggers a native double-free at the LiteRT layer).The most valuable Gemma 4 feature for a project like this was tool calling. Pre-Gemma-4 on-device LLMs would emit free-form JSON that I'd have to regex-and-state-machine my way through. With function calling, structured output is guaranteed — which is what made the offline study-material generation feel as reliable as a cloud product.
Sampling tuned for this app: temperature: 0.3, topK: 40, topP: 0.95
for tool calling (Gemma's published guidance for structured outputs — greedy decoding temp~0, topK=1
is prone to short repetition loops on long structured responses). The next time I walk onto a Cameroonian campus to talk about cloud and AI careers, I won't have to caveat the AI portion of the talk with "…of course you'll need stable internet and an English-fluent prompt." I can hand a student a Gemma-4-powered app, watch them point it at their own French-language course PDF, and watch them get a quiz in French, generated on-device, with no data plan involved. That's the bar for AI tools that actually work for the next billion learners — and that's the bar Gemma 4 cleared. Built with Flutter, Dart, flutter_gemma , ObjectBox, LiteRT-LM, and a healthy disregard for cloud dependencies. Follow me on LinkedIn or check out EduCloud Academy for what comes next.