This is a submission for the Gemma 4 Challenge: Build with Gemma 4
What I Built
The Spark Behind OptiLearn
Education in crisis-affected areas faces massive, systemic hurdles. Refugee camps often suffer from a severe lack of internet connectivity, scarce educational materials, and overwhelmed teachers managing large classrooms with vastly different student backgrounds. Modern, cloud-based AI tools are virtually useless in these environments because they require stable internet connections and ongoing infrastructure costs that these schools simply cannot afford.
We realised that to truly democratise AI in education, we couldn't rely on the cloud. We needed to bring the AI directly to the edge.
What It Actually Does
OptiLearn solves the connectivity barrier by running Gemma 4 entirely locally. It acts as a highly capable, always-available teaching assistant that empowers educators to optimise their teaching without ever needing to go online.
Through OptiLearn, teachers in resource-constrained environments can:
- Personalised Multilingual Tutors: An Arabic-speaking student gets an Arabic tutor at a Grade 3 level; a Somali student gets a Somali tutor at a Grade 5 level—all happening simultaneously from the same local server.
- Live Class Translator: The teacher speaks, local Whisper transcribes it, and Gemma 4 translates it to every student's native language in real-time.
- Adaptive Quizzes: Questions automatically adjust to the student's level, tracking their mastery using Exponential Moving Average (EMA).
- Teacher Heatmaps: A colour-coded grid lets teachers instantly see every student's mastery across all topics, generating automatic alerts before a student falls behind.
- 100% Local: Zero data leaves the classroom. It's completely private and offline.
Ultimately, OptiLearn proves that cutting-edge AI can be leveraged for good, delivering better, more personalised learning experiences in the most disconnected parts of the world.
Architechture Diagrams: https://miro.com/app/board/uXjVHSbGyBQ=/?focusWidget=3458764672167765361&embedMode=view_only_without_ui&embedId=351779570391
Demo
Demo Video: https://drive.google.com/file/d/18AanaC6ifMvufp9F7wn-UT2ZXZwbMg94/view?usp=sharing
Website: https://opti-learn.com/
Code
GitHub Repository: https://github.com/Ilakiancs/OptiLearn
How I Used Gemma 4
We built OptiLearn around Gemma 4 E2B, running locally via Ollama.
The Constraint: 2015 Hardware & Zero Internet
If we wanted this to work in a refugee camp, the AI couldn't rely on cloud API calls, nor could it require a massive GPU rig. We needed a model that could run efficiently on a donated 2015-era laptop with just 8GB of RAM, while still being smart enough to act as a pedagogical tutor.
Why Gemma 4 E2B Specifically?
Unmatched Efficiency: E2B was the exact right fit. It allows us to process multiple student requests locally, loading in about 15 seconds on older hardware without a dedicated GPU.
Native Multilingual Support: Refugee classrooms are incredibly diverse. Gemma 4 natively supports over 100 languages. Out of the box, we could deploy tutors in Arabic, Somali, Amharic, Tamil, and Tigrinya without relying on external translation APIs.
Trauma-Aware Fine-Tuning: Because the model runs locally, we also built a QLoRA fine-tuning pipeline (via Unsloth) in our repository. This allows the model to be explicitly trained on trauma-aware pedagogy using over 11,500 training examples, ensuring the AI never says "wrong" or "failed" to vulnerable students.
(Note: We also built a "Circuit Breaker" architecture. If the camp does get reliable internet, the system seamlessly falls back to the Google Gemini API to use the massive Gemma 4 26B model for richer translations, before silently failing back to the local E2B model when the connection drops).
OptiLearn proves that cutting-edge AI doesn't just belong in data centers—it belongs at the edge, where it can actually change lives.
Team Name: Opti5 Labs
Team Members: @ilakian (Ilakian Puvanendra), myself (Sasiru Virajith)
Top comments (0)