CAIRN: Enabling Cleaner AI Research

Team

Hongjin Lin, Computer Science PhD

Riad El Soufi, Master Design Engineering

Role

Product Design, User Research, Design System

Duration

4 weeks

Domains

AI, Sustainability, Energy

Deliverable

User Research

Key Outcome

Figma Prototype

Created design system and interactive Figma prototype for evidence-based product recommendation

16 Interviews

Conducted 16 interviews and translated user needs to product requirements

Validated Concept

The Assistant Director of the Harvard Office of Sustainability (primary user), expressed a need for our proposed concept

Training AI models are carbon intensive, and data center construction is accelerating rapidly

Training LLMs like GPT3 produces about 500 metric tons of CO2, equivalent to taking 500 flights from NY to SF.

We found the focus on individual AI researchers is misplaced

No researchers we interviewed took steps to reduce their carbon emissions. Multiple researchers expressed they want to adopt sustainable measures, but are unable due to the pressure to publish rapidly.

A renewable energy source has the greatest impact for decarbonizing AI research

We learned that the carbon emissions of AI research heavily depends on the data center location. Harvard's cleanest data center, MGHPCC, is on a grid powered by a hydroelectric dam. An opportunity exists to reduce emissions at the institutional level.

How might we shift focus to institutions from individual AI researchers?

Through interviews with the High Performance Computing Office and Office for Sustainability at Harvard, we came to understand the transition to the hydro-electric powered MGHPCC, completed in 2013. The Office for Sustainability led the effort. Due to their ability to implement institution-wide policies, we saw first hand this would be the way to drive results for other institutions making a similar transition.

We hypothesized that an impact assessment tool could be effective to decarbonize data centers

We brainstormed and prioritized potential solutions at the institutional level. We determined that an impact assessment tool for sustainability managers at universities, national labs, and corporations could be most effective.

Impact assessment is a slow multi-stakeholder process

We interviewed the Associate Director and a Sustainability Manager in the Harvard Office for Sustainability. We learned that data is essential to get buy-in for decarbonization projects. This is done with an impact assessment, which outlines the intended effects of a proposed policy.

Stakeholder management involved with data gathering is the main obstacle to accelerating the impact assessment process

We worked with the Harvard Office for Sustainability to understand their process of impact assessment. Data-driven goal-setting is essential to getting approval for sustainability policies. There is an opportunity to reduce the time associated with this slow and manual process.

Validated features with rough prototype

We created a rough prototype and showed it to our connections at the Office for Sustainability. This allowed us to validate our design decisions and gain a deeper understanding of their needs for 'guided learning.'

In IT impact assessments, jargon is a pain point

While interviewing our users about our prototype, they provided this significant caveat to their need for a guided data collection process.

4 Key Value Propositions

Users expressed a need for:
‍

Understanding and applying universal standards and guidelines
Guided learning for metrics, data sources, and institutional jargon
Communicating with multiple stakeholders in the data-collection process, across departments
Baselining with benchmarks

Introducing CAIRN

We created an IT emissions assessment builder for sustainability managers and their stakeholders. Most sustainability managers create impact assessments from their desktop computers, so we optimized for a desktop breakpoint. Our goal is to create a tool to overcome the most time consuming and manual processes of creating a decarbonized data center. We target a user who is aware of the problem, actively trying to find solutions to data center carbon emissions, and has pre-existing incentive structures.

Assessment builder

Guided form to collect data, following universal standards
Data collection can be assigned to stakeholders for accountability
Clear distinction between completed and uncompleted data fields

Stakeholder management

Side panel allows assigning data collection, commenting, and stakeholder tagging
Sustainability managers can share with stakeholders who receive notifications by email
Stakeholders access though emailed link and do not need an account

AI Helper

Side panel allows chat with AI Helper trained on IT industry standards and terminology
Utilize Retrieval-Augmented Generation to minimize hallucinations

Visualize & compare

Smart filters allow for fast report generation
Compare data across multiple computing centers, and with industry benchmarks
Buttons to save, export, print, and share reports

Key user flow: assigning data entry to an IT stakeholder

Cairn is intended to be seamless for collaborating with stakeholders within large institutions. For example, when an IT stakeholder is assigned data entry by a sustainability manager, the IT stakeholder is sent a notification email with a link to open the assessment directly, no need to create a username or password.

"This tool would be very useful to us within the next six months."

Associate Director, Harvard University Office for Sustainability

Colors, buttons, typography, and icons

I chose colors appropriate for a sustainability tool, both functional and warm. The button design is chunky and friendly, offering a satisfying tactility when pressed. The typography is intended to be clear and direct. Icons are from the extensive Google Material Design library.

Components created and organized in Figma

To ensure consistency, I started with the simplest components first, moving to increasingly more complex components. I strive to keep my Figma file highly organized.

In this CS276 course project I was able to make many new friends, develop my research skills, and deepen my understanding of human-computer interaction.