Democratisation of AI

Team name: Disruptors

Team members: Akshat Mehta (O.P. Jindal Global University), Manmohan Hooda (O.P. Jindal Global University), N. Prasanth (O.P. Jindal Global University), Daksh Dahiya (Maharshi Dayanand University)

Use of AI tools:

Yes, we used AI tools during the development of our submission.

Which tool(s) did you use?

Claude (Anthropic).

For what purpose(s)?

Research and drafting.

How did these tools support your work?

Claude helped us research open-source model benchmarks, hardware specifications, Indian CSR and education-policy provisions, and country-level cost estimates that we then independently verified. It also supported the drafting process turning our notes, decisions, and design choices into clean, structured prose across both the EduAI Global Implementation Report and the AI-Proof Assessment Platform Proposal. All ideas, design decisions, cost models, and the MDU pilot plan originated with the team; Claude accelerated the research and drafting cycle so a four-person student team could produce two long-form, internally consistent documents in a compressed timeframe.

External feedback & contributions:

Bruno DE LIEVRE Reviewer on the AI Grand Challenge platform Detailed written feedback on our initial contribution. He pushed us to clarify the link between our two core concerns (AI dependency weakening critical thinking, and unequal access deepening inequality) and to show how our solution addresses both, not only access. He also asked us to make our evidence more explicit costs, open-source tools, server needs, languages, user volumes and to propose a realistic pilot with one institution, a limited number of students, governance rules, data protection, teacher guidance, and evaluation criteria covering learning, equity, and responsible use. He recommended restructuring the argument to start with inequality, then move to infrastructure and anti-dependency.

Fatou Bintou Ba Reviewer on the AI Grand Challenge platform Written feedback affirming the framing of AI as shared infrastructure rather than an individual subscription, and noting that this shifts the conversation from access to ownership. She raised the critical follow-up question of how an open-access model can still encourage critical thinking rather than reproducing the over-reliance we set out to address.

Initial contribution: The Effect of Use of AI.

Final contribution:

Our final contribution is a two-part, integrated proposal addressing both the access problem and the integrity problem that AI creates in education:

EduAI Global Universal Open-Source AI for Every Student. A globally scalable framework for deploying open-source AI on-premises at every school and university. Each institution buys its own server once, runs open-source models (Llama 4 Maverick/Scout, Qwen2.5-Math, Qwen2.5-Coder, OpenBioLLM, DeepSeek), and routes students to department-specific models through an API gateway. Includes a teacher-controlled RAG memory system so every class has its own syllabus, lecture notes, and assessment rules. The MDU Rohtak pilot demonstrates 158+ crore in savings over four years versus cloud subscriptions, with full data sovereignty.

[Name Not Decided Yet] AI-Proof Assessment Repository. A free, open-source companion platform where faculty across every discipline build, test, and refine assessment methods that AI cannot defeat. Community-governed, AI-assisted, multilingual, and organised by academic discipline. It includes cross-disciplinary tactic matching, loophole reporting with AI recalibration, multi-sector validation tagging, and a community-sourced library of detection techniques.

Together, EduAI Global addresses access and dependency (every student gets capable AI, owned by the institution), and the assessment platform addresses integrity and critical thinking (faculty have validated methods so AI becomes a learning tool, not a shortcut). Neither solution is sufficient alone.

Drive folder with full documents and video pitch: https://drive.google.com/drive/folders/1ui9e6KTuz9UYg7bni8M-WK8A51bnn0_y

Reflection on the process how our contribution evolved during Phase 2:

Our Phase 1 submission focused primarily on the hardware-and-access angle: making capable AI available to every student equitably through institution-owned open-source infrastructure. The feedback we received on the platform in Phase 2 reshaped the proposal in important ways.

What feedback influenced our work:

Bruno DE LIEVRE's feedback was the most structurally consequential. He identified that our initial contribution raised two distinct concerns AI dependency weakening critical thinking, and unequal access deepening inequality but our solution only really addressed the second one. He pushed us to explain how the solution addresses both, not only access. He also asked us to make our evidence concrete: specific costs, specific open-source tools, specific server needs, specific languages, specific user volumes. And he asked for a realistic pilot one institution, a defined number of students, clear costs, governance rules, data protection, teacher guidance, and explicit evaluation criteria covering learning, equity, and responsible use.

Fatou Bintou Ba's feedback affirmed the framing of AI as shared infrastructure rather than subscription, and crucially raised the question of whether open access alone could simply reproduce the over-reliance we were trying to address that giving every student capable AI does not, by itself, build critical thinking.

Both comments converged on the same gap: solving access without solving assessment integrity would not solve the cognitive-development problem we identified in Phase 1.

What changes we made as a result:

We added the companion platform proposal ([Name Not Decided Yet] AI-Proof Assessment Repository) as a structural counterpart to EduAI Global. The two now form a single integrated thesis: institutional ownership of AI (hardware) + community-owned assessment knowledge (software/social). This directly answers Bruno's point that the solution must address both dependency and access, and Fatou's question about how open access avoids reproducing over-reliance.

We made the evidence explicit, as Bruno requested. The final report includes specific hardware specifications (Llama 4 Maverick FP8 with 2 and 4 H100 configurations), concrete cost lines (6071 lakh for Tier 1, 1.21.45 crore for Tier 2, 2.83.5 crore for MDU), a software stack listed component by component (vLLM, LiteLLM, Open WebUI, Qdrant, nomic-embed-text, Keycloak, Grafana), language coverage logic, and user-volume math for every tier.

We built out the MDU Rohtak pilot in full as the realistic single-institution pilot Bruno asked for: 14,00020,000 students, 59 departments, department-by-department AI model assignment, 68 H100 dual-node configuration, governance via a four-level admin hierarchy (Institution Admin Department Head Teacher Student), data sovereignty on-premises, teacher control via the RAG memory upload system, and a 4-year cost-vs-cloud comparison as an evaluation criterion.

We expanded the economic modelling from a single-institution pilot to a full 185-country, income-stratified, 4-year rollout budget (~$108B globally vs $1.2T/year in equivalent subscriptions) answering Bruno's question on whether the model is sustainable beyond one school.

We added the CSR and government co-funding model (50% State / 30% Centre / 20% institution) so the proposal is fundable, not just technically sound.

How this process strengthened our final proposal:

The biggest shift, prompted directly by Bruno DE LIEVRE's feedback, was moving from "AI for every student" to "AI infrastructure that institutions own, plus assessment infrastructure that faculty own." Phase 1 solved access. Phase 2 solves access, integrity, dependency, and equity simultaneously and does so through institutional design rather than restrictions on individual student behaviour. Fatou Bintou Ba's question about critical thinking gave us the clearest test for whether the proposal was complete: an access-only solution would have failed that test. The two-part structure passes it because faculty, not the AI vendor and not the student, hold the lever that determines whether AI becomes a tutor or a shortcut.

Video pitch:

https://drive.google.com/drive/folders/1ui9e6KTuz9UYg7bni8M-WK8A51bnn0_y

(The video is in the Drive folder above. If the form needs a direct link to the video file specifically, open the folder, right-click the video, "Get link," and paste that here.)