Imagine a high-tech classroom with no internet. That's the Mukono Digital Classroom, an early-stage experiment exploring what offline AI could look like in a primary school in rural Uganda. The prototype you're looking at is real and functional: speech recognition, translation between English and five Ugandan languages, text-to-speech in multiple voices, exam prep with automated grading, educational videos with translated captions and textbooks with translation at a click across all core subjects—all running on low-cost hardware.

The project is centered on Mukono Primary School, a public school in the Kunungu District of Uganda supported by the Pursuit of Excellence (POE) Scholarship Foundation. POE funds lunch, breakfast, teacher salaries and training, scholastic materials and provides financial scholarships for top performing students. On the national PLE exam, the school went from a total of 3 students scoring in the top division over the past decade to 11 in the most recent cycle, making it the second-best public school in the district. POE also built 5 new classrooms and a multi-purpose room with electricity where the Digital Classroom will be hosted.

Making this work on an 8 GB server when the best available translation models have never seen many academic terms takes creative problem-solving. The system simplifies concepts before translating, draws on custom glossaries, and uses fine-tuned models trained on Ugandan curriculum content. The result is an offline LLM that can explain topics, grade exam answers, and hold a conversation—all on low-cost hardware. Translation is powered by Sunbird AI's models fine-tuned for Ugandan languages, educational videos come from Kolibri by Learning Equality, and when connectivity is available the system can tap into Claude by Anthropic for richer, more detailed responses.

This is a working prototype with room to grow, and it would benefit hugely from collaboration—whether that's curriculum content, language expertise, funding, or help deploying hardware in schools. If any of that sounds interesting, or you just have feedback or questions, please reach out to Matt Goldman at [email protected].

Hardware & Cost

2× NVIDIA Jetson Orin Nano Super ($260 each), 256 GB NVMe SSD (~$25), 2× 64 GB microSD (~$8 each), 30× Samsung Galaxy Tab A9 (~$150 each), 30× wired headsets with mic (~$11 each), 30-port USB charging station (~$60), WiFi router (~$60), various cables and surge protectors (~$50). Total ~$5,600 per classroom.

Languages

Runyankole, Luganda, Acholi, Ateso, and Lugbara via Sunbird AI’s fine-tuned NLLB translation models. Additional local languages can be supported by wrapping around any suitable translation model.

AI Pipeline

Speech recognition (Sunbird Whisper Large V3 SALT), pronunciation checking (Whisper Tiny.en), translation (Sunbird NLLB 1.3B), LLM (Qwen 3.5-4B), English TTS (Piper), Runyankole TTS (VITS). All models run locally—no cloud required.

Architecture

Two-Jetson (fully offline): Jetson A (gateway) serves the UI, runs the LLM and English text-to-speech. Jetson B (worker) runs speech recognition, translation, and Runyankole text-to-speech. Everything works without internet.

Single-Jetson alternative (~$300): One Jetson runs translation, speech recognition, and text-to-speech fully offline. Q&A and explanations use the Claude API when internet is available.

30 student tablets connect via browser—no app install needed.

Content Safety

Extensively tested safety classifier ensures the model only engages with age-appropriate questions for East African primary school students. This classifier can be tuned for local norms.

Electricity

Each Jetson draws 7–15 W under load (~10 W average). A WiFi router adds ~8 W. Each Samsung Galaxy Tab A9 uses ~3.5 W during active screen-on use and needs ~19 Wh to fully recharge its battery. For a 6-hour school day:

• 2 Jetsons: 2 × 10 W × 6 h = 120 Wh
• WiFi router: 8 W × 6 h = 48 Wh
• 30 tablets (recharge): 30 × 19 Wh = ~570 Wh
• Total: ~740 Wh (0.74 kWh) per day

That’s less than a single lightbulb left on for a day. Peak wall draw is under 300 W (even with all 30 tablets charging simultaneously through a USB charging cart), well within a single 240 V outlet. A small UPS or solar-battery setup could keep the system running during outages.

Connectivity

Fully offline by default. Optional Claude integration when internet is available. The tablets have no SIM cards, no cellular data, and no internet access—they connect only to the classroom WiFi network. Away from the Jetsons they are little more than basic Android tablets with no content, making them low-value theft targets.

Using Translation & Simple Explanation to Understand Complex Academic Words in Local Languages

Assigning Educational Videos to Students and Tracking Their Progress

Select a class above to view student progress, or create a new class.

NCDC Syllabus

Official Uganda curriculum topics with links to textbook pages

◀ Prev Next ▶ Close