AI-VERDE: A Gateway for Egalitarian Access to Large Language Model-Based Resources For Educational Institutions

Edwin Skidmore; Enrique Noriega-Atala; Nirav Merchant; Paul Mithun

REVIEW 3 major objections 6 minor 1 cited by

AI-VERDE: A Gateway for Egalitarian Access to Large Language Model-Based Resources For Educational Institutions

T0 review · 3 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A university-run LLM gateway, tested for six months, handled 97,658 API calls and 110.7 million tokens for 78 users, and the paper argues this shows a practical route to campus-wide equitable access.

desk verdict Useful, honest systems paper about a University of Arizona LLM gateway; the cost comparison it promises never materializes, undercutting the central 'egalitarian' claim. read the letter →

arxiv 2502.09651 v1 pith:LJOA6BUS submitted 2025-02-11 cs.CL cs.CY

Paul Mithun , Enrique Noriega-Atala , Nirav Merchant , Edwin Skidmore This is my paper · ORCID

classification cs.CLcs.CY

keywords LLMplatformasaservicehighereducationretrieval-augmentedgenerationprivacy-preservingAIAPIaccessmanagementon-premisesmodelhostingcost-effectiveuniversitypilotstudy

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

AI-VERDE is a university-hosted gateway that lets instructors, students, and researchers reach both open-source and commercial language models through one login and one API key, with per-course budgets, access controls, and built-in retrieval-augmented generation. The paper claims this is the first platform designed to serve instructional and research needs together inside a higher-education institution, and that it removes the cost, privacy, and technical barriers that keep LLMs off campus. In a six-month pilot at a large public university, 78 users across five courses and ten research projects made 97,658 API calls worth 110.68 million tokens, mostly through self-hosted models. The sympathetic reading of the pilot is that even this modest voluntary adoption demonstrates a realistic route to campus-wide equitable access if hardware and staff support scale with demand.

What carries the argument

The carrying mechanism is a multi-tenant LLM gateway built around an OpenAI-compatible reverse-proxy API layer. Requests arrive at one access point and are routed to whatever model the caller names, whether a model hosted on institutional GPUs by a high-throughput serving engine or a commercial endpoint reached through a budget-controlled proxy. Around that core sit per-course vector databases for retrieval-augmented generation, a document-ingestion service, API-key and budget management, and single sign-on through the university's identity system. This stack converts the usual adoption problems, privacy, cost, authentication, model choice, and RAG setup, into configuration choices rather than per-user technical projects.

What would settle it

Total the hardware, power, staff, and software cost of the pilot's 110.68 million self-hosted tokens and compare that with commercial API pricing for the same volume; if self-hosting costs as much or more, the cost-egalitarian claim fails.

Watch

Extended reading notes

Core claim

The central claim is that one platform can unify all LLM access for an academic community and make that access egalitarian rather than dependent on personal subscriptions or technical skill. AI-VERDE is the proposed mechanism: each course or research group gets its own workspace, vector database, budget, and API keys, while all data stays on institutional infrastructure. The pilot numbers are offered as evidence: 97,658 API calls and 110.68 million tokens across 78 users, five courses, and ten research projects over six months. The paper reads this usage as "significant engagement" and "huge potential for adoption," while acknowledging the raw numbers are low for an institution of that scale.

Load-bearing premise

The load-bearing premise is that 78 voluntary pilot users and 372 survey respondents represent what the whole campus will need, and that the institution will continue to fund the hardware and human consultants as the user base grows.

Editorial extensions

If this is right

A course can get a spending cap, automatically generated API keys for every enrolled student, and instant revocation at the end of the term.
Sensitive course and research data can be used with LLMs inside institutional infrastructure, which is what makes privacy-regulated work feasible.
Because the API matches a widely used standard, existing tools and code assistants work without modifications.
One gateway can serve instruction, research, and support departments, replacing several separate pay-per-use arrangements.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Going beyond what the paper proves, the metering data could be used to decide where to invest in GPUs versus commercial credits, since 99 percent of the pilot's tokens were self-hosted.
An unmeasured variable is the human support bundled into the platform; comparing adoption across departments with and without AI-specialist consultants would separate the software's contribution from the service's.
A direct extension would compare learning outcomes from the sandboxed course chatbot against a general-purpose chatbot, putting the hallucination-reduction claim on empirical footing.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Dear colleague,

This paper is a straightforward systems description of AI-VERDE, an LLM gateway for universities that combines vLLM, LiteLLM, Weaviate, CILogon, and Kubernetes into a managed multi-tenant service with course-level access control and budget management. What is actually new is the integration: I don't know of another published account of a university-run platform that gives instructors automated API key provisioning, per-course RAG, and proxy access to commercial models with budget limits. The authors deployed it for real at the University of Arizona and report concrete usage numbers over six months: 78 users, 97,658 API calls, 110.68M tokens. That's credible engineering evidence of adoption, and the architecture section is clear enough for another institution to replicate.

The paper does several things well. The feature comparison table is honest, though it marks "Yes" for their own guardrails without detail. The discussion of institutional barriers—financial compliance, FERPA/HIPAA, onboarding overhead—reflects genuine engagement with campus IT realities. The appendix describes their survey honestly, including snowball sampling, which is a limitation they don't hide. The ethics statement is candid about proxy privacy limits and hardware constraints.

The soft spots are real, and the stress-test note is on target. The introduction promises a comparative cost study with commercial options, but none appears. There is no hardware cost, no personnel cost, no per-token or per-user cost, no baseline against OpenAI or Anvil. Table 2 shows 109.76M self-hosted tokens vs 0.919M proxy tokens, which suggests most traffic is on their own GPUs, but without fixed costs we cannot evaluate the "much lower cost" claim. This is load-bearing because the paper's central value proposition is egalitarian access, and egalitarian access rests on affordability. Second, the "first of its kind" claim is asserted, not verified by a literature search. Third, claims about privacy and reduced hallucination are plausible but unmeasured; the privacy claim is partially walked back in the ethics section. These are gaps a competent referee could push them to fix, not fatal flaws.

Verdict: This is a useful infrastructure case study, not a research discovery. For a systems or demo track, it deserves a serious referee, but it needs major revision: add cost data, soften or substantiate the "first" claim, and report any security or accuracy evaluation. I'd bring it to a reading group only if the conversation is about LLM platform deployment in academia.

Recommendation: send to peer review, conditional on major revisions.

Referee Report

3 major / 6 minor

Summary. The paper presents AI-VERDE, an LLM platform-as-a-service deployed at the University of Arizona that unifies access to self-hosted open models (served via vLLM), commercial model APIs and research services such as AnvilGPT behind a LiteLLM proxy, and adds course/group management, per-course budget controls, native RAG support through a managed Weaviate instance, and a conversational web interface with CILogon authentication. The authors report a six-month pilot with 78 users across 5 courses and 10 research projects, totaling 97,658 API calls and 110.68 million tokens, and they argue from a 372-participant qualitative survey that the platform addresses privacy, intellectual property, cost, and usability barriers that block LLM adoption in higher education. The paper claims that AI-VERDE is the first platform to address both instructional and research LLM needs within a higher education institutional framework and that it provides a lower-cost 'egalitarian gateway' relative to commercial options.

Significance. If the cost and adoption claims were substantiated with data, this systems contribution would be directly useful to university computing centers and comparable institutional service providers. The architecture description is concrete and replicable, the deployment statistics are internally consistent, and Table 2 usefully disaggregates self-hosted versus proxy token volumes and API calls. The authors' explicit acknowledgments of limitations (dedicated hardware constraints, the privacy boundary for proxied commercial traffic, and residual model bias) are commendable and improve the paper's credibility. However, the central value proposition of an egalitarian, lower-cost gateway is asserted rather than demonstrated: no cost comparison is provided, the pilot population is small and self-selected, and strong claims about reduced hallucination and HIPAA-grade privacy are made without measurement. As it stands, the paper is best read as a deployment blueprint and requirements catalog, not as an evaluated solution to the affordability problem it highlights.

major comments (3)

[Section 1, Contribution 3; Section 3.4; Section 3.6; Appendix F] The paper promises a comparative study against commercial options showing a 'much lower cost egalitarian gateway' (Section 1, Contribution 3), and Section 3.4 claims 'minimal operational costs' through open-source components and partnerships, yet no cost comparison appears anywhere in the manuscript. Section 3.6 and Appendix F contain only the statement that hosting LLMs often exceeds $50,000; there are no figures for AI-VERDE's hardware, personnel, per-token, or per-user costs, and no baseline against OpenAI, Gemini, or Anvil pricing. Table 2 shows that 109.76M of the 110.68M tokens were self-hosted, which shifts cost from per-token fees to fixed infrastructure expense, but that fixed expense is never disclosed. Because the egalitarian-access claim rests on affordability, this missing cost evidence is load-bearing and must be closed with actual deployment costs or a clearly scoped cost model.
[Section 3.5 and Appendix B.2] The claims of 'HIPAA-grade' privacy via the Soteria integration and of 'minimal hallucination' via a specially engineered prompt are made without any audit, compliance certification, or evaluation. No HIPAA compliance assessment is cited, and the hallucination discussion provides no test set, baseline model, or quantitative result such that a reader could verify the claim. These claims should either be removed or replaced with measurable evidence, and the paper should state what certification or review the Soteria integration has actually undergone.
[Section 4.2 and Appendix A] The generalization that AI-VERDE has 'huge potential for adoption in academic settings' is not supported by the evidence presented. The pilot involved 78 users, a number the authors themselves describe as 'relatively low' for an institution of this scale, and the survey in Appendix A recruited through sign-up flyers, snowball referrals, and an open campus call, so the 372 participants cannot be treated as representative of the university community. The paper should either reframe the conclusions as pilot-scale observations or provide a sampling plan, response rates, and demographic coverage information that would support population-level claims about egalitarian access.

minor comments (6)

[Section 4.2] The text says 'more than 110 tokens were passed' but the table to which it refers reports 110.68 million tokens; the 'million' is missing.
[Abstract and Section 1] The phrase 'an higher education institutional framework' should be 'a higher education institutional framework'; similar article errors appear elsewhere (e.g., 'an higher education institution' in the abstract).
[Appendix C] Items 7 through 9 are all labeled 'Harwood Lab' but one of the URLs points to Joseph Bonito's page, so the labels do not match the cited URLs; the list also contains a duplicated label for Harwood across items 7 and 8.
[Appendix F] Appendix F is titled 'Cost details' but contains only the same single sentence as Section 3.6; either expand it with the actual cost analysis promised in the introduction or rename it to avoid implying that cost data are provided.
[References and Section 2.1] The text cites Touvron et al. (2023) for Llama 3.2, but that reference is the Llama 2 paper; the Llama 3 models are covered by Dubey et al. (2024), so the citations should be aligned with the models actually served.
[Appendix B.4] The example prompt contains empty <Reference></Reference> tags, which appears to be a rendering artifact and should either be filled in or removed. In addition, the phrase 'hoi polloi' in Section 1 is too informal for a journal venue.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; central claims are architecture descriptions and measured pilot usage, not fitted predictions.

full rationale

AI-VERDE is a systems paper with no fitted parameters, equations, or predictive derivations. The central claims are that the platform provides certain features and that a six-month pilot showed engagement; the usage statistics in Table 2 are measurements of the deployed system, not quantities derived from inputs by construction. The survey in Appendix A motivates the problem statement but does not define or entail the reported outcomes, and no result is justified by citing the authors' own prior work as an external theorem. The self-citations to CyVerse and Jetstream2 appear only as infrastructure acknowledgments and are not load-bearing for the paper's claims. The skeptic's concern about the absent cost comparison is a legitimate correctness/evidence gap, but it is not a circularity: failing to support a claim is different from making the claim true by definition. Accordingly, no circular step can be exhibited with the required specificity, and the appropriate score is 0.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

This is a systems paper with no mathematical derivations or fitted parameters. The load-bearing assumptions are all domain-level: survey representativeness, privacy guarantees, and the interpretation of usage statistics. No new scientific entities are postulated.

assumptions (3)

domain assumption The survey sample of 372 self-selected participants is representative of the university community's AI needs.
Section 3 and Appendix A describe the survey but do not establish statistical representativeness. Recruitment was voluntary and used snowball referrals, which likely introduces selection bias.
domain assumption On-premises hosting fully eliminates privacy and compliance risks.
Sections 3.1 and 3.5 assert privacy guarantees without a security audit. The paper itself concedes in the Ethics and Limitations section that proxy usage to external platforms cannot carry the same guarantee.
domain assumption Usage volume, measured in API calls and tokens, is a valid proxy for educational and research value.
Section 4.2 treats token counts as evidence of adoption and potential, but no outcome measures such as learning gains or research productivity are provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AI-VERDE: A Gateway for Egalitarian Access to Large Language Model-Based Resources For Educational Institutions." pith.science (2026). https://pith.science/paper/LJOA6BUS

@misc{pith2026250209651,
  author       = {Pith},
  title        = {Pith review of: AI-VERDE: A Gateway for Egalitarian Access to Large Language Model-Based Resources For Educational Institutions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LJOA6BUS}},
  note         = {Machine review of arXiv:2502.09651}
}

read the original abstract

We present AI-VERDE, a unified LLM-as-a-platform service designed to facilitate seamless integration of commercial, cloud-hosted, and on-premise open LLMs in academic settings. AI-VERDE streamlines access management for instructional and research groups by providing features such as robust access control, privacy-preserving mechanisms, native Retrieval-Augmented Generation (RAG) support, budget management for third-party LLM services, and both a conversational web interface and API access. In a pilot deployment at a large public university, AI-VERDE demonstrated significant engagement across diverse educational and research groups, enabling activities that would typically require substantial budgets for commercial LLM services with limited user and team management capabilities. To the best of our knowledge, AI-Verde is the first platform to address both academic and research needs for LLMs within an higher education institutional framework.

Figures

Figures reproduced from arXiv: 2502.09651 by the authors.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

LLM-HyPZ: Hardware Vulnerability Discovery using an LLM-Assisted Hybrid Platform for Zero-Shot Knowledge Extraction and Refinement
cs.CR 2025-08 conditional novelty 5.0 of 10

A zero-shot LLM classification, embedding, and clustering pipeline labels 1,742 hardware vulnerabilities in the 2021-2024 CVE corpus and feeds MITRE's MIHW 2025 list.

Reference graph

Works this paper leans on

17 extracted references · 13 canonical work pages · cited by 1 Pith paper

[1]

Course content for RNR355: Introduction to Wildland Fire

work page
[2]

arXiv preprint arXiv:2408.13247

Data exposure from llm apps: An in- depth investigation of openai’s gpts. arXiv preprint arXiv:2408.13247. Albert Q Jiang, Alexandre Sablayrolles, Arthur Men- sch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guil- laume Lample, Lucile Saulnier, et al. 2023a. Mistral 7b. arXiv preprint arXiv:2310.06825. Albe...

arXiv
[3]

Content from website and documentation for CyVerse: A computational framework de- signed to handle large datasets and complex analyses

work page
[4]

Content from website and documentation for Tech Launch Arizona 27: Facilitating the com- mercialization of University of Arizona inven- tions

work page
[5]

Content from website and documentation MK- Docs 28: A static site generator for documen- tation projects

work page
[6]

Research publications from the University of Arizona’s Cooperative Extension 26

work page
[7]

Content and publications from Harwood Lab" 30

work page
[8]

Content and publications from Harwood Lab" 31

work page

Show all 17 references

[9]

Content and publications from Harwood Lab" 32
[10]

Content from website and documentation MK- Docs 29: an AI-powered Civil Engineering Community
[11]

Content and publications related to Antenna research for Hao Xin lab: 34 26https://extension.arizona.edu/ 27https://techlaunch.arizona.edu/ 28https://www.mkdocs.org/ 29https://www.chishiki-ai.org/ 30https://comm.arizona.edu/person/jake-harwood 31https://comm.arizona.edu/person...
[12]

Provides access to AnvilGPT models 35
[13]

AI is not a done deal. We are building the road as we walk it, and we can collectively decide what direction we want to go in, together

Course content for INFO 523 2024 Fall : Data Mining and Discovery. D Some more potential and future use cases In this section we detail some more future work, which are either planned or on which already de- velopment is being done as a feature addition to AI-VERDE . D.1 Visib...

2024
[14]

Content and publications from Eller Partner- ship Office: 33
[2016]

That question is beyond my purview of current knowledge

Learning management systems (lms) and e- learning management: an integrative review and re- search agenda. JISTEM-Journal of Information Sys- tems and Technology Management, 13(2):157–180. OpenAI. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774. TB OpenAI. 2022. ...

2023 arXiv
[2024]

Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al

The faiss library. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783. David Y Hancock, Jeremy Fischer, Joh...

2024 arXiv
[9474]

Shervin Minaee, Tomas Mikolov, Narjes Nikzad, Meysam Chenaghlu, Richard Socher, Xavier Am- atriain, and Jianfeng Gao

Curran Associates, Inc. Shervin Minaee, Tomas Mikolov, Narjes Nikzad, Meysam Chenaghlu, Richard Socher, Xavier Am- atriain, and Jianfeng Gao. 2024. Large language models: A survey. arXiv preprint arXiv:2402.06196. 7 Paulo Cristiano de Oliveira, Cristiano Jose Castro de Almeida...

2024 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Course content for RNR355: Introduction to Wildland Fire

work page

[2] [2]

arXiv preprint arXiv:2408.13247

Data exposure from llm apps: An in- depth investigation of openai’s gpts. arXiv preprint arXiv:2408.13247. Albert Q Jiang, Alexandre Sablayrolles, Arthur Men- sch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guil- laume Lample, Lucile Saulnier, et al. 2023a. Mistral 7b. arXiv preprint arXiv:2310.06825. Albe...

arXiv

[3] [3]

Content from website and documentation for CyVerse: A computational framework de- signed to handle large datasets and complex analyses

work page

[4] [4]

Content from website and documentation for Tech Launch Arizona 27: Facilitating the com- mercialization of University of Arizona inven- tions

work page

[5] [5]

Content from website and documentation MK- Docs 28: A static site generator for documen- tation projects

work page

[6] [6]

Research publications from the University of Arizona’s Cooperative Extension 26

work page

[7] [7]

Content and publications from Harwood Lab" 30

work page

[8] [8]

Content and publications from Harwood Lab" 31

work page