OKAE — Open Kit for AI in Education: A Reference Architecture for Offline, Curriculum-Grounded AI Tutoring on the Raspberry Pi 5

OKAE (Open Kit for AI in Education) is an open-source reference architecture for personalized, curriculum-grounded, multilingual AI tutoring that runs entirely on a Raspberry Pi 5 with 8 GB of RAM. It is built for the 2.2 billion children and young people the internet cannot reach: 27 percent of the world's population, two thirds of school-age children in low-income countries, predominantly non-anglophone, in schools without IT infrastructure, recurring software budgets, or per-student devices.

This deposit publishes v4.3 of the architecture as a coordinated release. Seven files are attached: the canonical Whitepaper, the Technical Specification, the Privacy & Data Posture document, an executive One-Pager, the source-code tarball of the canonical OKAE release, and the Apache 2.0 LICENSE + NOTICE. All artifacts are released under Apache License 2.0.

This is a restricted-access deposit. The metadata, abstract, author, keywords, and DOI are public and citable. The seven attached files require an access request to download — see the access conditions on the deposit page. Researchers, ministries of education, deployment partners, and academic collaborators are warmly invited to request access.

What v4.3 ships

* On-device tutor LLM: IBM Granite 3.3 2B (Apache 2.0), ~1.5 GB resident, model-swappable via a single environment variable.
* Curriculum-grounded retrieval: pre-built ChromaDB vector index over 916,215 chunks across 2,724 sources from seven content providers — NCERT (India), OpenStax, CK-12, Kiwix, Project Gutenberg, NCERT Exemplar, and local uploads — with per-source attribution and citation discipline.
* Seven-layer K-12 safety toolkit: a deterministic educational-only gate, system-prompt safety rules, input sanitization, output filtering, a multilingual panic-phrase listener (1,044 phrases across 8 languages), an append-only audit log with erasure-time redaction, and a teacher dashboard with per-student export and delete.
* Personalization layer: per-student profile and PIN-based identity (no OS account required, no third-party account of any kind), mastery tracking with a spaced-repetition guard against single-session cramming, optional teacher notes, optional weekly objective, multiple learning-style modes including Socratic dialogue, and a `#fast` per-turn opt-out for quick general-knowledge questions.
* Multimodal: text, voice (faster-whisper + Piper TTS), and optional handwriting evaluation via Granite Vision 3.2 2B with a Tesseract OCR fallback.
* Multilingual: translation modes (off / glossing / full-L1) across the reference model's twelve native languages, plus a regex-based panic listener that covers an additional safety-critical language set independent of model language coverage.

Why a Raspberry Pi 5

The Pi 5 is the smallest, cheapest, most repairable, most procurement-defensible piece of commodity hardware on which a full-featured K-12 AI tutor can run with no internet, no vendor account, and no recurring fee. The architecture is built around shared classroom use: one Pi serves a rotation of students through one or two station endpoints (a display + USB keyboard, connected over the Pi's own Wi-Fi hotspot or HDMI cable). Base hardware — the Pi 5, its required Active Cooler, power supply, and a 128 GB card sized to the full published curriculum — lands in the $100–$200 range depending on region and on what a school already has.

One Pi draws approximately 8 W under sustained inference — about the power of a single LED bulb. An entire classroom Pi cluster averages 25–40 W for the school day, structurally four to seven times more power-efficient per student than one-device-per-student deployments. A modest solar add-on sustains the cluster off-grid indefinitely — opening the architecture to rural schools, refugee settlements, NGO field deployments, and disaster-resilience scenarios.

Every benchmark figure in the attached documents was measured on the shipping hardware and is reproducible with a one-line command against the open-source harness: 98.8% non-failure and zero hallucinations on a 162-question K-12 bench, 100% pass on a 41-question adversarial bench.

What's in this deposit

* `1_OKAE_Whitepaper_v4.3.pdf` — the v4.3 Whitepaper: architectural thesis, access-deficit analysis, model selection rationale, structural cost analysis.
* `2_OKAE_Technical_Spec_v4.3.pdf` — the v4.3 Technical Specification: RAM and storage accounting, RAG store design, the seven-layer safety toolkit, model-selection benchmark methodology, per-turn token budget, deployment topology, ingestion pipeline, runtime configuration.
* `3_OKAE_Privacy_v4.3.pdf` — the v4.3 Privacy & Data Posture document, including the encryption posture (SQLCipher at rest, per-device key, passphrase-derived backups) and retention limits.
* `4_OKAE_One_Pager_v4.3.pdf` — executive one-pager.
* `5_okae-v4.3-source.tar.gz` — pristine snapshot of the canonical OKAE codebase at the tagged v4.3 release.
* `6_LICENSE.txt` + `7_NOTICE.txt` — Apache 2.0 license text and the attribution NOTICE required under §4(d).

Reusability

The Apache 2.0 license is permissive: approved requesters may use, deploy, build commercial services around, and adapt the architecture without royalties or revenue-sharing with the original author. Deployment partners price their own classroom installation, support, training, and curriculum customization services separately — those costs are outside the scope of this deposit and this architecture.

**Citation**

Please cite as: Srivats, B. G. (2026). *OKAE — Open Kit for AI in Education: A Reference Architecture for Offline, Curriculum-Grounded AI Tutoring on the Raspberry Pi 5* (v4.3) [Software]. Zenodo. doi.org

Author: Bharath G. Srivats · github.com

 

This is a restricted-access deposit. The published metadata (title, abstract, author, keywords, DOI) is public and citable. The seven attached files — Whitepaper, Technical Specification, Privacy & Data Posture, executive One-Pager, the source code tarball of the canonical OKAE release, plus LICENSE and NOTICE — are gated behind author approval. See the access conditions field below.

Researchers, educators, ministries of education, deployment partners, and academic collaborators are invited to request access via Zenodo's built-in access-request flow or by emailing the author directly.

ACCESS CONDITIONS

Access to the attached files is granted at the author's discretion to verified deployment partners, academic collaborators, ministries of education, and NGO operators who can articulate a specific use of OKAE in service of K-12 students who lack reliable internet connectivity.

To request access, please include in your request: (1) your name and affiliation, (2) a one-paragraph description of the deployment context or research question you intend to apply OKAE to, (3) confirmation that you accept the terms of the Apache License 2.0 under which OKAE is published.

Requests are typically reviewed within one to two weeks. Approved requesters receive a time-limited download link via Zenodo's notification system.

For urgent partnership inquiries (school districts in active deployment planning, time-sensitive grant submissions, etc.), email the author directly at bharathsrivats@outlook.com.