LLAMA 3.1 70B

Major Research University

CASE STUDY

Helping higher education share knowledge and work smarter with a Llama-powered GPT solution

At a glance

Industry: Education

Use case: Streamlining information discovery and automating repetitive tasks

Goal: Increase productivity for 37,000 faculty, staff and student employees

Llama versions: Llama 2 7B, 3.1 405B and 3.3 70B

Deployment: Open-source pipeline running on-premises in Major Research University’s supercomputing center

Download full case study

62%

reduction in time spent writing job descriptions

87%

reduction in time spent searching HR, finance and facilities policies and documents

80%

reduction in time spent analyzing and redlining contracts

*All results are self-reported and not identifiably repeatable. Generally expected individual results will differ.

THEIR STORY

Leading in higher education and tech innovation

As one of the world’s top-20 research facilities, Major Research University (MRU) is a public institution with more than 43,000 students and 37,000 employees offering degrees in the arts and sciences, engineering, business, medicine and more. The school is a hub for technological innovation home to numerous startups, a data center institute and a world-class supercomputing center.

THEIR GOAL

Improve information access and streamline tasks for staff and faculty

MRU had a vast knowledge base of institutional information — including policies, procedures, contracts and technical resources — across dozens of departments, institutes and schools. MRU’s IT team wanted to build a GPT solution that could make that information far more accessible and usable via intuitive chat agents in a secure environment. Ultimately, they envisioned helping staff, faculty and student employees save time and work smarter.

THEIR SOLUTION

Llama powers a suite of university AI assistants and coaches

In a 12-month sprint, MRU’s IT team used Llama to build a multi-agent fleet of AI assistants backed by university knowledge bases. The agents help with everything from internet searches and job descriptions to fund management and preventing email phishing.

MRU chose Llama as the solution’s primary large language model (LLM) for its performance, cost savings, transparency and safety features. Llama performed better than competing open and closed models for accuracy, F1 scores, precision and recall. As an open-source solution, Llama allowed MRU to bypass the licensing fees and per-token charges that come with commercial models. Llama also gave developers more fine-tuning control and the ability to innovate quickly, upgrading from Llama 2 to Llama 3 as new models were released.

THEIR APPROACH

Multiple security layers ensure MRU stays in control of its data

To protect privacy, sensitive university data must never leave the school’s domain or be touched by a third-party service. With Llama, the GPT solution can be hosted entirely on premises at the school’s supercomputing center to ensure control.

MRU’s IT team took data security a step further with the help of Protopia and their Stained Glass Transform (SGT) technology. SGT uses AI to transform data into randomized re-representations that an LLM can consume without exposing the source. The combination of on-premises infrastructure and SGT gives the solution the security required to index sensitive data and generate answers with minimal risk of data leaks or breaches.

The Llama-powered MRU GPT solution uses an agentic approach that prioritizes open-source software whenever possible.

THEIR SUCCESS

Significant productivity gains across the university

With Llama-backed AI agents, MRU has eased information gathering and automated routine and complex tasks. The solution does more than save time for staff and faculty, it lessens friction throughout the university.

Projected results:

• 62% reduction in time spent writing job descriptions

• 87% increase in first-time-right chip manufacturing

• 80% reduction in time spent analyzing and redlining contracts

*All results are self-reported and not identifiably repeatable. Generally, expected individual results will differ.