Offline-first, distraction-free educational AI for K-12 iPads.
All inference runs locally on the Apple Neural Engine. No internet. No surveillance. No distractions.
School districts across the United States spend billions deploying 1:1 iPad programs. The intent is learning. The outcome is TikTok. The device meant to replace textbooks has instead replaced attention spans.
Existing solutions — content filters, MDM lockdowns, screen time limits — treat the symptom. They do not treat the cause. A student who cannot access TikTok will find the next distraction. The deeper problem is that the device is inherently networked, and networks are inherently adversarial to focus.
The only real solution is to sever the network entirely.
Alexander OS is an offline-first educational operating environment that runs on existing school iPads. It does three things no existing edtech product does simultaneously:
- Severs the internet at the OS level during study sessions — not via filtering, but via true network isolation. The device cannot phone home, stream, or be surveilled.
- Runs Small Language Models entirely on-device using the Apple Neural Engine and the MLX framework. No API calls. No cloud. No latency. No data leaving the device.
- Acts as a Socratic tutor, not a search engine. Instead of giving students answers, it directs them back to the exact page in their physical textbook — rebuilding reading habits and deep focus.
The internet is severed at the session level. Alexander OS operates in a true offline mode — no background syncs, no telemetry, no API calls. Student queries never leave the device. This is not a content filter; it is architectural separation.
Running two language models simultaneously on a base iPad (4–6GB RAM) causes thermal throttling and memory crashes. Alexander OS solves this with a lightweight subject-detection kernel that hot-swaps between two quantized models:
| Model | Parameters | Use Case |
|---|---|---|
Llama 3.2 |
1B (4-bit quantized) | Humanities, History, Literature, General |
Qwen 2.5 Math |
1.5B (4-bit quantized) | Mathematics, Physics, Chemistry, STEM |
Only one model is loaded into the Neural Engine's memory pool at any time. Switching takes under 2 seconds on Apple Silicon.
Alexander OS is explicitly prohibited from giving direct answers. Its system prompt enforces a pedagogical constraint: every response must reference the student's physical textbook by chapter and page, ask a guiding question, and build toward understanding rather than completion. It is designed to make the student think, not to think for the student.
Because all inference is local, there is no user data to expose. No query logs are transmitted. No student profiles are built. No third-party APIs process student input. Compliance with FERPA, COPPA, and state-level student privacy laws is structural — not contractual.
| Layer | Technology |
|---|---|
| On-device inference | Apple MLX |
| Hardware acceleration | Apple Neural Engine (ANE) |
| STEM model | Qwen 2.5 Math 1.5B (4-bit via MLX) |
| Humanities model | Llama 3.2 1B (4-bit via MLX) |
| Development environment | macOS, Python 3.11+ |
| UI prototype | SwiftUI (iPadOS 17+) |
| Quantization | mlx_lm.convert with --q-bits 4 |
| Fine-tuning | MLX LoRA fine-tuning pipeline |
alexander-os/
│
├── README.md ← You are here
│
├── ui-prototype/ ← SwiftUI mockups and screen designs
│ ├── screens/ ← Individual screen layouts
│ └── assets/ ← Icons, color palette, typography
│
├── experiments/ ← Core logic prototypes (Python)
│ ├── alexander_logic.py ← SLM kernel: subject detection + hot-swap
│ ├── rag_pipeline.py ← Retrieval from school-uploaded PDFs
│ └── socratic_prompt.py ← Prompt engineering for Socratic mode
│
└── docs/ ← Technical documentation
├── architecture.md ← Edge-AI system design
├── hardware_specs.md ← Development hardware requirements
└── privacy_model.md ← Student data protection framework
This repository represents the pre-alpha research and architecture phase.
- Core SLM hot-swap kernel (Python prototype)
- Subject detection logic
- Socratic prompt architecture
- System architecture documentation
- Hardware specification documentation
- MLX model quantization pipeline
- SwiftUI prototype (in progress)
- On-device RAG from school PDFs
- MDM/school deployment configuration
"The most dangerous thing you can do to a child's mind is give them infinite content and infinite answers and call it education."
The global edtech industry is optimised for engagement. Alexander OS is optimised for the opposite: focused, deep, offline learning. It is the first educational AI tool designed to make itself less necessary over time — by redirecting students to books, not to screens.
Varun Jarwani — Ahmedabad, India
18-year-old engineer and founder. Built the core logic prototype while preparing for India's Joint Entrance Examination (JEE).
MIT License. See LICENSE for details.