Magis
Product development
nmillrr
Role:
Research engineer
Date:
Jan 2024
Preview
Coming soon
Making education more interactive and accessible
As a senior thesis project, I spent a semester researching, developing, and designing a deployment strategy for a humane tech product. I decided to take a look at education in rural, low-resource areas that could benefit from better access to information.
I took inspiration from YouTube videos, Reddit, and Instructables that focused on doomsday/offgrid internet devices. These projects helped me get a better sense of what basic materials and programs a device should be loaded with, what types of power generation/storage I could use, and how to build for different environments (rain, heat, cold, etc.).
We're so obsessed with making new and novel solutions in tech—it's easy to forget that low-tech solutions can often work perfectly for modern problems, even at scale. Below, you can read about my final product, Magis, and the impact we were able to have.
Information accessibility in low-resource areas
While a majority of modern technological investment is focusing on cutting-edge developments in AI, much of the world still hasn’t established strong internet infrastructure. This is what’s known as the digital divide: innovation in the Western world is fixated on superintelligence and AGI, ignoring foundational accessibility issues that under-resourced classes and regions are continuing to face. Around the world, two-thirds of school-age children have no home internet, leaving about 1.3 billion students without online resources. Many remote and rural communities–like mountain villages, refugee camps or underserved regions–lack reliable power and connectivity. As a result, students often rely on outdated textbooks, radio broadcasts, or infrequent mobile signals for basic educational information. In 2020, school closures due to the COVID-19 pandemic highlighted how problematic the digital divide has become: in low-income regions, only a tiny minority of homes had internet access, which led to severe learning gaps and disruption.
Different low-resource areas like Latin America, central Africa, and rural Asia face different challenges local to their region, such as language differences and limited digital skills. These challenges exacerbate educational inequity, as native-language instruction and context-aware support are often unavailable. In areas of conflict, like Sudan or Palestine, or frequent natural disasters, like the Philippines, physical networks are often down, isolating entire communities from information.
Optimally, investment in strong internet infrastructure would be a sustainable long-term solution for such regions. But with limited resources, countries and communities might want to opt for offline learning systems that are able to operate without the internet. This would allow citizens to access information at low costs, without continuous power supply, and even in harsh physical environments. This is how we can make education accessible for everyone, even where the internet never reaches.
Existing offline solutions
Internet-in-a-Box (IIAB): IIAB is a proven offline “learning hotspot” used in dozens of countries. It typically runs on a Raspberry Pi or similar device, broadcasting a local Wi-Fi that serves a curated library of content. Standard IIAB distributions include Wikipedia (in many languages), Khan Academy videos, OpenStreetMap, e-books and more. Users nearby can connect with phones, tablets or laptops and browse this “digital library” without any internet access. For example, communities in the Dominican Republic have deployed IIAB boxes to share medical guides, lecture videos, and textbooks.
Complementary initiatives: Other projects like SolarSPELL and RACHEL likewise deliver offline libraries (Wikipedia, videos, health guides) via local servers. These systems often bundle content packs for education and health. Here’s a video of how the SolarSPELL library is used in rural clinics and schools, sometimes powered by solar energy.
While IIAB and RACHEL improve access to static content, they lack interactive search and summarization. Students must navigate menus or conduct keyword searches through content packs. They cannot ask nuanced questions (“Explain Newton’s second law in simple terms”) or get customized explanations. There is no built-in ability to summarize long articles or translate on demand. In essence, existing IIAB systems are like offline encyclopedias–informative, but non-interactive. Additionally, without the internet, these devices can’t fetch new information or answer questions beyond the pre-loaded material. Students often need guidance (e.g. tutoring or clarification) that static content cannot provide.
Beekee, a similar system, is compact and solar-operable, providing a full “digital classroom” offline—but for an $800 price tag.
Proposed AI + IIAB solution
The base solution will involve augmenting the Raspberry Pi-based IIAB with a lightweight on-device small language model (SLM) called TinyLlama (1.1B)—to enable offline natural-language Q&A and summarization using the existing IIAB library. The system would work as follows:
Hardware: The recently released Raspberry Pi 5 (or similar) with 8 GB RAM, powered by a small battery or solar panel. It hosts the IIAB software and Wi-Fi hotspot, plus the AI model. No external GPU or cloud connection is needed—TinyLlama is small enough to run on the Pi’s CPU.
Software: We install the standard IIAB image (with chosen content packs: Wikipedia, Khan Academy, MedlinePlus, TEDx transcripts, etc.) and the Ollama framework (or similar) for local LLM inference. TinyLlama is loaded into Ollama. We also install a lightweight retrieval engine (e.g. using embedded text search or a vector database like Qdrant on-device) that indexes the local content.
AI Model (TinyLlama): TinyLlama is a 1.1 B-parameter open-source SLM (architecture based on LLaMA-2). We chose it because, when quantized to 4 bits, it runs with under ~1 GB RAM and responds in <1 second on a Pi 5. This model can generate humanlike answers or summaries, and it can be fine-tuned or prompted for conversational use. We will integrate retrieval-augmented generation (RAG): before each query, the Pi’s retriever fetches relevant passages from the local content (Wikipedia articles, Khan Academy notes, etc.), and includes them in the prompt. This grounds TinyLlama’s responses in actual offline data, improving accuracy and reducing hallucinations.
User Interface: Users connect to the IIAB Wi-Fi (e.g. “SunnyBox”) and open a web browser to the local portal. The portal features a chat-like Q&A interface: learners type questions in their language and receive answers synthesized by TinyLlama. They can also request summaries of lengthy chapters or video transcripts, all done instantly by the on-device SLM. Optionally, voice input/output could be added for students with low literacy.
Benefits: This enhancement turns IIAB from a passive library into an interactive tutor. Students can ask open-ended questions (e.g. “What causes rainbows? Explain simply.”) and get on-the-spot explanations. Teachers can request summaries of complex topics to prepare lessons. Medical workers can query health guidelines (from MedlinePlus or WHO content) in natural language. Importantly, everything works offline: once TinyLlama and the content packs are installed, the Pi requires no internet or cloud. Updates could be pushed occasionally via USB drives or short network access, but day-to-day use is self-contained.
Framework for Global Deployment
While a lightweight Raspberry Pi might work as an MVP in most parts of the world, some countries and communities have different levels of access to electricity, connectivity devices, and information. We propose a tiered framework for different regions by levels of resource access, with tailored solutions suited to each context.
Low Power, Low Connectivity: Countries with frequent power outages, little to no grid access, weak or nonexistent internet, and minimal local devices, like rural parts of sub-Saharan Africa, need the lightest version of a system. A suitable product would be built on a lightweight Raspberry Pi, cheap I/O devices, solar-charged batteries, and a small language model (SLM) rather than an LLM. This approach would minimize energy consumption and required technical resources, and still delivery interactive content offline. Multiple stations might be set up in community centers and classrooms such that the system could be accessed by all.
Power Available, Few Devices: Countries with reliable electricity but limited hardware might be better equipped for a higher-tech solution. Low-income communities in regions like the Cuban countryside or rural Latin America might benefit from such a system. Hardware could be upgraded from communal stations to individual, refurbished rugged laptops. The computers could run stronger LLMs (TinyLlama up to 2-3B in size) and host the full IIAB content library, and would still be able to run off of battery packs in the case of outages. Students in these regions would get access to deeper content interactivity, more educational media, and even be able to take learning home with them.
Devices & Power, Low Information Access: There are a few countries and regions where both device ownership and reliable electricity are common, but information access is not. Areas like Iran, Venezuela, and Turkmenistan struggle with poor internet infrastructure or selective censorship. In these communities, a Raspberry Pi or a stronger computer (like the BeeKee) could act as a local hotspot for individuals to connect to with their own devices. Hotspots are cheaper, more resilient, and more efficient for these regions compared to purchasing separate, new devices (as long as such devices are correctly configured to avoid remote monitoring in areas of censorship).
Commercial-grade laptops can be found refurbished or secondhand for pennies on the dollar. The above models retail for around $2,000, but when companies cycle out the older technology, it gets sold for much cheaper prices.
Why TinyLlama for the MVP
We explored SLM and LLM options for the most basic Magis product, and landed on TinyLlama for it's incredible size-to-performance ratio. At the time of writing this in early 2024, there is no other open-source model that is lighter than 1B parameters without failing to adequately fulfill our use case.
Size and Performance: TinyLlama (1.1 B parameters) is specifically designed for resource-limited devices. When quantized, it occupies only ~700 MB on disk and requires under 1 GB RAM on a Raspberry Pi 5. In practical tests, TinyLlama responses took well under 1 second on a Pi 5, enabling nearly real-time Q&A. Larger models (like Llama-2 7B) simply wouldn’t run on the same hardware. Thus TinyLlama is a practical choice for an on-device AI assistant.
Open-Source License: TinyLlama is released under the Apache-2.0 license, which is permissive and compatible with open-source projects like IIAB. Any modifications or integrations we build can be shared freely. Its architecture and tokenizer match LLaMA-2, meaning it can leverage the growing ecosystem of Llama-based tools and community models.
Community Support and Maturity: TinyLlama has active community development on Hugging Face (many downloads). It can be plugged into existing LLM frameworks and fine-tuned if needed. We benefit from an evolving ecosystem of quantization, optimization, and deployment tutorials. The model’s creators explicitly optimized it for “restricted computation and memory footprint”—exactly our use case.
Capabilities: Despite its small size, TinyLlama delivers impressively coherent responses for many educational tasks. In robotics research, it was used with on-device RAG to improve decision-making, confirming that retrieval-augmented TinyLlama can “achieve more precise, context-aware” outputs. We expect similarly improved accuracy when TinyLlama is fed relevant IIAB content context.
Prototype Implementation
Required Hardware
Raspberry Pi 4 or 5 (8GB RAM)
SD card
Power supply, power bank, rechargeable batteries, or solar power module
Wi-Fi adapter (only for initial setup)
Implementation
Installing IIAB content: Use the one-line IIAB installer or image to flash the Pi with IIAB’s latest release. Configure the content by selecting languages and packs (Wikipedia slices, Khan Academy videos, MedlinePlus, TED transcripts, etc.). Ensure the device’s Wi-Fi hotspot is active and functional.
Install Ollama and the SLM: Install the Ollama LLM runtime on Raspberry Pi OS (see Ollama docs). Use Ollama to download TinyLlama: e.g.
ollama pull tinyllama:latestSetting up the retrieval engine (RAG): Prepare the IIAB content for retrieval: either build a local full-text index of the HTML/PDF files, or convert key articles into a vector database. For example, run an indexing script over the selected Wikipedia articles and Khan transcripts on the SD card. Store the index files on the Pi’s SSD/SD for fast lookup. Implement a simple retriever (open-source libraries like whoosh or small-vector DB).
Integrating the chat interface: Develop or adapt a web-based chat UI that runs on the Pi (e.g. a Flask or Node.js app). The frontend sends the user’s query to a backend script. The backend first queries the RAG index for the top N relevant documents, then crafts a prompt including those passages and the user’s question, and calls TinyLlama (via Ollama CLI or Python API) to generate an answer. The answer is returned to the UI. Ensure the UI is language-flexible (supporting Spanish, indigenous languages, etc.) and mobile-friendly.
Testing and iterating: Populate the Pi with a complete set of test queries covering math, science, languages, and health. Verify that TinyLlama provides correct, context-grounded answers. Tune the RAG pipeline (e.g. number of retrieved docs, prompt formatting) to minimize hallucination. Collect feedback from educators or local pilots to refine the system.
Impact
Improved information access: By turning static content into an interactive tutor, learners in remote areas gain immediate explanations and personalized guidance. Instead of scrolling endlessly through articles, a student can simply ask, “How do I calculate area of a triangle?” and get a concise, step-by-step answer. This on-demand Q&A mimics a tutor or teacher — which is often unavailable in low-resource schools. Studies show that AI tutors can accelerate learning by providing individualized attention; doing this offline could dramatically boost outcomes in under-served communities.
Educational equity: Offline IIAB devices break down geographic and socioeconomic barriers. Communities without internet will no longer be permanently excluded from the AI revolution. As one German analysis notes, “SLMs (Small Language Models) offer a transformative solution for democratizing education” in areas lacking connectivity. This aligns with global equity goals: 4.5 billion people without internet are currently “left out” of AI benefits. Our system helps ensure that students in Latin America, Africa, or any region can ask questions and learn in their own language, closing the digital divide.
Resilience through disaster and crisis: Offline AI is inherently resilient to infrastructure failures. In emergencies (earthquakes, hurricanes, wars), when internet and power may be disrupted, a battery- or solar-powered Magis continues to operate. Magis could be packed into disaster relief kits to provide survivors with health instructions, emergency guides, and educational content – all without any network. In this way, learning and critical information remain available even in crises.
Language and content diversity: Because IIAB supports multiple languages (e.g. Spanish, Portuguese, indigenous languages), Magis can be configured to respond in the local language. TinyLlama’s architecture allows fine-tuning or prompting in any language present in the dataset. This is crucial for global applicability (Latin America, Africa, Asia, etc.) where English is not dominant. Learners can query the system in their native language and receive culturally relevant answers.
Scaling educational resources: The combination of IIAB’s curated content and TinyLlama’s generative abilities means each deployment becomes a full-featured offline learning center. Schools or NGOs can distribute Magis units to many villages at marginal cost, dramatically increasing reach. In pilot projects (like ASU’s SolarSPELL), students report that having an offline AI companion allows them to “find information more efficiently and effectively” in the local library. We expect similar improvements: teachers will spend less time searching for materials and more time teaching.
Future Applications
In the longer term, Magis could enable new community services beyond basic learning. Complex and more heavyweight systems like e-commerce, banking, and entertainment might require more serious digital infrastructural upgrades, but critical domains like health, education, and history can be scaled with incremental investment into offline technologies.
Health: Embedding medical knowledgebases (like MedlinePlus or local health manuals) into the RAG index allows Magis to act as a health advisor. Field nurses could ask clinical questions (e.g. “What are the symptoms of dengue fever?”) and receive vetted answers. This follows SolarSPELL’s model: their NextLab team fine-tuned an LLM on the SolarSPELL health library so that nurses could ask it questions offline. Our solution could similarly support rural clinics with 24/7 AI-driven consultation.
Furthering education: Magis can serve as a tutor in basic literacy, math, or vocational skills. For instance, adults could practice language learning, home improvement, or business acumen. Magis could also host quiz apps or interactive exercises generated by the AI. Importantly, running TinyLlama on-device (as in the ASU project “EDgeAI”) empowers learners with AI practice even when offline.
Agriculture: Magis could guide farmers on optimal agricultural techniques. Edge AI combined with sensors would allow for advancements in industrialized farming and agritech: a key component for accelerating a society's further development. At ASU, students used “TinyML” on a soil sensor to categorize moisture and send results offline to farmers. In our context, Magis could integrate local data (soil, weather forecasts downloaded in advance) and answer questions like “Will it rain tomorrow?” or “When should I plant corn?”. This would help farmers in low-resource regions adapt to climate variability.
Historical preservation: Over time, the system could incorporate historical knowledge (folk stories, indigenous science, traditional medicine) into an archival database. Magis can then help document and disseminate cultural knowledge, which is often threatened in low-resource areas.