AI Integration Report

Team Number: 7

Team Name: fooBaz

Introduction

Our OIM project is directed at making money management accessible for illiterate and

innumerate adults, and as a result, anything that may usually be communicated in text

must be represented with icons for our user interface. Hence, we needed to generate a

large library of icons which would be used in core features of our project such as showing

transactions, spending categories, and notications. The pending generation of these

generate consistent, quality icons that could be used for the features in our app. Without

this AI solution, we would have needed to rely on icons that were free for commercial use

online. This would generally result in lower-quality icons for our users that are less clear in

their messaging. In addition, there would be less cohesion between the icons since there

did not exist an icon set that had all the icons we required.

We measured success of our AI solution in four ways:

1. The time spent on each icon

2. The revision cycles required per icon (how many times we had to remake the icon or

for the AI, how many times we needed to reprompt it)

3. The consistency of the style for the icons

We evaluated multiple options for the AI generated icons we needed. Our main criteria

when looking for AI Tools here, was that it had to meet our needs for visual consistency,

control, and eciency. It also had to match the style of the preexisting Wallet OIM Icons, so

that it would not break the design style of the UI. Furthermore, the icons were being made

for a innumerate and illiterate audience, so the icons also could not be overly abstract or

artistic. Though the process was mostly smooth, there were some issues associated with

asset generation. Firstly, Gemini would quite frequently not understand the instructions, or

deviate from the style, which required us to rephrase and increase our iterations, to get it to

within a reasonable style. This was denitely a frustrating challenge to deal with. Another

challenge was that certain icons required multiple revision cycles. That is, we had to start

provided limited reproducibility. Next, we considered a directly integrated tool , which was

the Figma AI Plugin tool. We had experience using Figma, and this was integrated and

fast to use. However, it oered very little customization, and thus we could not control it

to produce acceptable icons. We nally decided to use Google Gemini after some test-

ing. We found that Gemini balanced control and creativity together the best. Additionally,

Gemini’s generated images were completely free for commercial use. Although it had a

weakness with stylization drift occurring, it was still the best tool we found for the task. Our

main workow for generating the icons was that we would pick a preexisting icon from the

codebase, and use it as a source of ground truth for Gemini to base its style o of. Then,

we would ask it to generate a a version with our desired changes to the icons, and we

Our AI-powered pipeline demonstrates signicant improvements in visual quality and

consistency across all icon categories. As shown in Appendix .1, the transformation from

legacy to AI-generateds exhibits enhanced clarity, addition of colours, and a more rened

styling with respect to older legacy code. The pipeline successfully enhances edge de-

nition and overall visual coherence while maintaining the core semantic meaning of each

icon. Across all comparisons, the AI-generated versions display superior line quality, aes-

thetic consistency, and a unied visual language that contributes to a more polished and

professional user interface. The complete icon gallery in Appendix ?? showcases both

the evolved icons and those created directly through the AI renement process. The con-

sistent improvement across these diverse icon types validates the eectiveness of our

user experience and brand consistency.

Impact Analysis

Because our users are illiterate or innumerate, icons are the primary way people under-

stand actions in the interface. Clear, consistent visuals were essential, so improving our

asset workow directly improved usability. Without AI, each icon would have required 120-

150 minutes of manual work, including sketching, rening, testing variations, incorporating

feedback, and exporting nal assets. For 21 icons, this would have totaled roughly a full

week of focused design time. Using AI, the average time per icon dropped to 12.4 minutes.

Prompting, selecting variations, and making light edits still required supervision, but all 21

icons were completed in under 5 hours. This resulted in an eciency gain of about [(150

from various AI tools were often too abstract or overly detailed. We had to re-prompt

frequently and reset context due to free-plan limitations. Several icons still needed manual

simplication or resizing in Figma. While AI accelerated creation, it still required design

judgment to control noise and ensure readability on mobile screens. Without AI, we would

have spent a full week designing icons manually or purchased licensed icon packs costing

≈ $20. Hiring a designer would have been even more expensive ranging from $100 -

$1000 per icon. AI tools introduced no monetary cost, and the only overhead was time

spent iterating on prompts and cleaning certain outputs. Gemini’s royalty-free license also

allowed free use and modication of generated assets.

Overall, the benets signicantly outweighed the costs. AI improved speed, consis-

To begin, each team member explored a dierent AI model to evaluate its potential

for generating usable assets. Nathan focused on Google Gemini, Anish tested DALL·E

3, Zaki explored Midjourney v6, Fares experimented with Ideogram, and Amish worked

with Figma’s native AI plug-ins. We compared models based on style consistency, prompt

responsiveness, export quality, and ease of renement. After several iterations, Gemini

clearly outperformed the others. Its multimodal interface allowed both text and image

prompt renement, and it maintained cohesive icon sets across multiple generations. The

ability to generate icons in batches with balanced proportions and stylistic uniformity made

Gemini the most reliable option for integration into our workow.

To measure Gemini’s impact, Amish conducted a controlled experiment in Figma. He

convenience.

Our collaboration emphasized clear structure and iterative renement. Nathan han-

dled experimentation and optimization, ne-tuning settings for batch generation and eval-

uating image-to-prompt feedback loops. Anish led prompt design and established a cen-

tralized zero-shot chain-of-thought prompt system to standardize how visual inputs were

structured. This approach improved reproducibility and visual consistency across outputs.

Zaki developed an automated post-processing script to convert Gemini outputs into scal-

able vector formats for integration with our front-end pipeline. Fares conducted aesthetic

and accessibility reviews to ensure color contrast, line clarity, and recognizability across

devices. Amish contributed to benchmark testing and documentation of manual versus AI

cycles, particularly for scalable UI components and marketing visuals. For other teams, we

recommend experimenting with multiple models before committing, documenting prompt

structures, and maintaining human oversight during renement. When used thoughtfully,

AI can act as a genuine creative collaborator, enhancing productivity while preserving hu-

man intent and design quality.

.1 Gemini Prompts

1. Prompt 1

2. Prompt 2

3. Prompt 3

4. Prompt 4