Teacher-Led, Localised AI for Equitable Education | Global Roundtable

Contents

Executive Summary

The Commonwealth of Learning is launching a compact for "frugal AI" in inclusive education—a distributed, locally-hosted approach to AI that operates offline, preserves data sovereignty, and reduces computational costs. The initiative, already piloted in five countries across India, Africa, and the Pacific, positions teachers as central curators of AI-generated content aligned with government-approved curricula, addressing educational access challenges in remote and resource-constrained regions.

Key Takeaways

Frugal AI proves that advanced AI benefits are not exclusive to well-resourced institutions — Decentralized, locally-hosted systems can serve remote communities, small nations, and under-resourced schools without sacrificing pedagogical quality or data security.
Teachers must be active participants, not passive recipients — The most critical differentiator from commercial AI is the teacher-in-the-loop architecture. Educators retain authority over what content enters classrooms, preserving instructional quality and cultural/linguistic appropriateness.
The Global South is producing the innovations needed for the Global South — India's frugal AI framework, developed by IIT Jodhpur and deployed via Commonwealth of Learning, directly addresses constraints in African, Pacific, and other developing contexts without imposing Western technology stacks.
Offline, low-bandwidth deployment is not a limitation—it's a liberation — Institutions are freed from cloud vendor lock-in, monthly bills, internet dependency, and data privacy risks. For regions with intermittent connectivity, local-first architecture is essential, not optional.
Policy frameworks must precede technology rollout — Countries piloting frugal AI report success correlates with clear leadership commitment, inclusive policy-making (involving all stakeholders), and active enforcement—not with technical features alone.

Key Topics Covered

Frugal AI architecture — local server deployment versus cloud-based systems
Teacher-in-the-loop design — educators as content curators and quality gatekeepers
Data sovereignty and privacy — keeping data within institutional/national boundaries
Reach to underserved regions — remote, rural, and small island nation contexts
Skills and vocational education — bridging technical education gaps in the Global South
Policy frameworks for AI adoption — leadership buy-in, policy development, enforcement
Curriculum alignment — training models on government-approved curricula
Multilingual and cultural adaptation — addressing linguistic diversity and local contexts
Sustainability and energy efficiency — reducing environmental footprint versus large language models
Addressing the digital divide — equitable access across urban/rural and wealthy/resource-poor divides

Key Points & Insights

Frugal AI differs fundamentally from commercial systems — It runs locally without cloud dependency, requires minimal bandwidth (works on 3G), uses no GPUs, and gives full control of data to institutions. This eliminates privacy concerns and external costs while enabling offline functionality.
Teacher agency is built into the system design — Unlike ChatGPT, the system requires teacher verification at every stage: content generation → verification → review → publication. Teachers curate outputs against government curricula before deployment to learners, maintaining pedagogical integrity.
Current deployment demonstrates feasibility at scale — The system is already operational in five countries (India, Ghana, Zimbabwe, Kenya, Nigeria) across school, college, and vocational training levels, with applications spanning mathematics, biology, fashion technology, and lab sciences.
Open universities in India face acute scalability challenges — Institutions like BAOU serve 300,000+ students across diverse languages, with student-to-faculty ratios that make personalized support impossible without automation. Frugal AI reduces content creation duplication and enables customization in multiple languages.
Small island nations face unique structural constraints — Pacific island countries have populations under 1 million, experts wearing multiple roles, limited IT infrastructure, and infrequent connectivity. Frugal AI eliminates external dependence and allows regional knowledge repositories to remain under local control.
Competency-based assessment requires different AI approaches than knowledge transfer — Kenya's experience shows that genuine skills development requires AI systems focused on products (what students create) rather than knowledge recitation. This reframes how AI evaluation and feedback should function in technical/vocational education.
Policy precedes implementation — Success requires leadership buy-in, clear policies developed inclusively (with students, faculty, administrators), and enforcement mechanisms before deployment. Policy-first approach prevents ad-hoc adoption.
Cost and sustainability arguments carry weight in fiscally-constrained contexts — For island economies and developing nations, large language models' energy demands, cooling requirements, and data hosting costs are prohibitive. Frugal AI delivers equivalent functionality at 10-20% of the resource cost.
Data sovereignty is both a technical and geopolitical necessity — Keeping education data within national/institutional boundaries protects against loss of control, aligns with sovereignty concerns, and prevents reliance on foreign infrastructure during geopolitical tensions or natural disasters.
Quality does not trade off against affordability — Curriculum alignment and teacher-in-the-loop design ensure output quality by design, not through post-hoc filtering. The system produces varied responses to the same prompt, but all remain aligned to approved curricula.

Notable Quotes or Statements

"Frugal AI means we have our AI server running in a local system. It's not on cloud. That gives some inherent advantage of offline ready... data sovereignty, you have full ownership of data because the data is not leaving your premises." — Dr. Sumit Kalra (IIT Jodhpur)

"We are talking about teacher in the loop system, which is very very important... the teacher is there in the loop for the conversation and solving the problem—that's really very important, especially in rural India where connectivity is a huge problem." — Professor Ami Upadia (BAOU, Vice Chancellor)

"You cannot transform students without transforming the teacher. We usually say: you fix the teacher, the teacher will fix the student, the student will fix the product." — Dr. Edwin Tano (Kenya School of Technical and Vocational Training)

"If I asked you to show how to make a stool, ChatGPT will give you an answer in 10 seconds. But ChatGPT hasn't made the stool. We're giving you 5 marks for showing how, and 95 marks for making the stool." — Dr. Edwin Tano (illustrating competency-based education)

"The island will cease to exist. For us, frugal AI is very important because it uses less power, less cooling system, so less energy, less carbon footprint." — Dr. Sokut Rohit (Special Adviser to Minister, Mauritius, on climate/sustainability concerns)

"When we talk about global south, this is the south I'm representing... small island countries with populations under a million, some just 2 meters above ocean level. Their identity, knowledge, culture, and tradition should not be lost." — Dr. Rajni Chand (University of the South Pacific, representing 13 Pacific island nations)

"We have to enforce it. You can enforce this using several strategies across the systems in your institution... that is the most difficult part." — Dr. Edwin Tano (on implementation challenges)

Speakers & Organizations Mentioned

Primary Speakers:

Dr. Sumit Kalra — Associate Professor, Department of Computer Science & Engineering, Indian Institute of Technology (IIT) Jodhpur
Professor Kabirun (First name unclear) — Indian academic discussing open universities and GER (Gross Enrollment Ratio) targets
Professor Ami Upadia — Vice Chancellor, BAOU (Baou Gujarat Open University) since 2018
Dr. Edwin Tano — CEO, Kenya School of Technical and Vocational Training (TVET)
Dr. Sokut Rohit — Special Adviser to the Minister for Tertiary Education and Research, Mauritius
Dr. Rajni Chand — Director, Center for Flexible Learning, University of the South Pacific, Fiji
Peter Scott — President and CEO, Commonwealth of Learning (intergovernmental organization of 56 Commonwealth nations)

Organizations/Institutions:

Commonwealth of Learning — Launching the "Compact for Frugal AI for Inclusive Education"
IIT Jodhpur (Indian Institute of Technology Jodhpur) — Developing the frugal AI system
BAOU (Gujarat Open University) — Serving 300,000+ students in India
University of the South Pacific — Regional university serving 13 Pacific island countries + Papua New Guinea
Kenya School of Technical and Vocational Training (TVET) — Leading competency-based education reform
Ministry of Education, Mauritius — Policy-level engagement
IIT Madras — Mentioned as conducting parallel summit on school education

Countries/Regions Referenced in Deployment:

India (5+ countries deployment)
Ghana
Zimbabwe
Kenya
Nigeria
Mauritius
Fiji and 12 other Pacific island nations (Tuvalu, Kiribati, etc.)
Papua New Guinea

Technical Concepts & Resources

Core Architecture & Concepts:

Frugal AI — AI systems optimized for resource-constrained contexts: local hosting, low bandwidth, minimal computational requirements, no cloud dependency
Teacher-in-the-loop — Architecture requiring educator review/curation at every content generation stage, not just final output
Curriculum alignment — Training language models on government-approved curricula rather than general internet data
Small Language Models (SLMs) — Contrasted with large language models (LLMs); domain-specific, smaller parameter count, deployable locally
Data sovereignty — Keeping training and operational data within institutional/national boundaries

Deployment Requirements:

Local server architecture — Institutional data center deployment (not cloud)
Bandwidth profile — Compatible with 3G connectivity and intermittent internet
Mobile-friendly design — Accessible via smartphones, not just laptops
No GPU dependency — Runs on standard server-class machines
Offline functionality — Core features work without internet connectivity

AI Models & Frameworks:

Open-source models — Referenced but not specifically named (speaker mentions "50 models" tested for different subjects)
Deep Seek — Mentioned briefly in Q&A as an example of open-source alternative
Generative AI agents — Used for content structuring, verification, and quality alignment
Subject-specific model variants — Separate models trained for mathematics, biology, fashion technology, lab sciences

Pedagogical Concepts:

Competency-based education — Focus on demonstrated skills/products vs. knowledge recitation
Personalized learning — Localized adaptation rather than one-size-fits-all approaches
Open Educational Resources (OER) — Reusable, sharable curriculum content
Choice-based credit system — Mentioned in context of National Education Policy (India)
Skill development frameworks — Addressing employability gaps (India: 51% employability, up from 33%)

Policy & Implementation Frameworks:

Leadership buy-in — Critical first step before deployment
Inclusive policy development — Involving students, faculty, administrators, and faculty heads
Enforcement mechanisms — Strategies for sustained adoption post-implementation
Competency verification protocols — For teacher training and competency assessment

Quality Assurance Mechanisms:

Multi-stage review process — Generation → verification → review → publication
Curriculum-bounded output — Models trained to stay within approved curricula even when prompted to deviate
Educator content curation — Final authority retained by teachers before deployment to students

Sustainability Metrics:

Energy consumption — Large language models: high cooling, high electricity; frugal AI: minimal
Carbon footprint — Critical concern for small island nations
Total cost of ownership — Cloud-based vs. locally-hosted comparison

Contextual Notes

This roundtable represents a policy announcement (Commonwealth of Learning's frugal AI compact) rather than a research presentation. The emphasis is on deployment experience in existing pilots across five countries and policy rationales rather than technical benchmarks or comparative performance data. The transcript quality degrades significantly in the Q&A section, suggesting audio processing challenges in the source recording.