Back to all works

Revamping the AI Model Development Life Cycle

AIMMO End-to-End Solution Tool Redesign

Demo Link

↗

Demo Link

↗

Solution

DURATION

3 Months

CLIENT

AIMMO

ROLE

Lead Product Designer

YEAR

2023

About Company

AIMMO: AI Data Solutions

Industry

AI/Data Solution

Headquarters

Seongnam, South Korea

Company size

201-500+

AIMMO provides advanced AI-driven data solutions across industries such as autonomous driving, smart cities, robotics, drones, and construction. By combining human expertise with AI, AIMMO enhances efficiency, accelerates innovation, and delivers real-world applications that optimize technology use at scale.

Top Achievements

Market Innovation

3 Months

to deliver to 0 → 1 MVP

Delivered the V1 End-to-End platform in time for the critical CES Global Launch, enabling the sales team to demo to VIP stakeholders.

• Met a critical 3-month deadline to secure a successful demonstration for VIPs at the CES launch

• Established a foundation for a new, unified product line, shifting the company's focus from siloed annotation tools to a comprehensive deployment platform.

Delivered the V1 End-to-End platform in time for the critical CES Global Launch, enabling the sales team to demo to VIP stakeholders.

• Met a critical 3-month deadline to secure a successful demonstration for VIPs at the CES launch

• Established a foundation for a new, unified product line, shifting the company's focus from siloed annotation tools to a comprehensive deployment platform.

Operational Efficiency

+60

%

Internal User Satisfaction

Systematically resolved critical UX debt across core labeling and model training workflows. Implemented efficiency-focused UI improvements (collapsible sidebar, clearer error states) which drastically reduced task completion time, leading to a validated 60% boost in internal user satisfaction.

Background

Why a core revamp was essential?

Challenge: From Annotation Tool to All-in-One Model Deployment Platform

• AIMMO had a working annotation tool but no full end-to-end platform

• Market research showed no all-in-one solutions existed, despite customer demand

• Collaborated with the business team to validate insights through interviews

• Tasked with designing a prototype version under a tight CES deadline for demonstration

• AIMMO had a working annotation tool but no full end-to-end platform

• Market research showed no all-in-one solutions existed, despite customer demand

• Collaborated with the business team to validate insights through interviews

• Tasked with designing a prototype version under a tight CES deadline for demonstration

Company focused exclusively on data annotation.

The goal is to enhance the product lines to offer comprehensive end-to-end model training and deployment.

Objective

The main objective was to streamline processes, minimizing labor intensity and boosting user engagement, in line with the company's Q4 goals for operational optimization.

UX Research

Research & Discovery: Phase 1

Company Seeks Comprehensive All-in-One AI Model Deployment Service

Before we started the redesign, we surveyed our customers and worked with the buisness team to collect customer testimonials. The company's input was essential in finalizing the next year deal.

“Essential Elements for Upcoming Contracts?”

MITSUBISHI

Emphasized that uploading large files for driver training increases operational load due to scalability challenges.

MAGNA

It was highlighted that server scalability directly increases operational burden.

PIXELPLUS

Emphasized that uploading large files for driver training increases operational load due to scalability challenges.

OUR HOME (CBM)

CBM Data: No automation of logistics volume-related measurements.

Research & Discovery: Phase 2

How can we boost productivity for users of our annotation tools, both internal and external?

We conducted field interviews with 20 internal users to identify UX issues, leading us to initiate significant UX improvements.

Interview insight:

65%

of Group A users

Admins & Project Managers

"It's hard to scan project progress, everything's buried in a long, scrollable list"

Interview insight:

80%

of Group B users

Labeling Workers

"I wish the sidebar would collapse or shrink when working with lots of content"

Research & Discovery: Phase 3

To design the tool, I sought feedback from AI developers and the MLOps team.

The development team and I met at least twice a week to design the product, as this is a tool for AI developers. To effectively tackle the project, I created a research plan to identify and address knowledge gaps, enabling informed design decisions. Understanding the AI training tool issues and evaluating competitors' offerings were crucial for the successful redesign of the product.

Raw file to model deployment simple version of information architecture

Raw Files

(Data from Autonomous Vehicles)

Dataset List

Data curation for

high-quality model training

Using artificial data

(Synthetic data)

to improve the training of models

Model

Workflow > Training > Deployment

Research & Discovery: Phase 4

Reviewed the end-to-end flow of Model deployment

I reviewed the end-to-end flow to gain a better understanding of the model deployment process.
I created the flowchart to identify which touch points caused pain points and to model the deployment flow from scratch.

Solutions

Comprehensive AI Training Product Redesign

Solution 1

Enhanced Labeling System

Improved the labeling process by integrating an automated failure history log. Users can now download detailed Excel or CSV reports of any upload failures, facilitating quick issue resolution and ensuring data integrity.

BEFORE

AFTER

Frequent failed GT uploads and ambiguous data quantities were difficult to identify.

Data uploads are critical to project progress.

Interviewed internal labelers to prioritize and address the most significant UX issues first.

If the failure reason is long, it should be downloadable.

For optimal user interaction, the "Retry all" button should have low affordance.

BEFORE

AFTER

Frequent failed GT uploads and ambiguous data quantities were difficult to identify.

Data uploads are critical to project progress.

Interviewed internal labelers to prioritize and address the most significant UX issues first.

If the failure reason is long, it should be downloadable.

For optimal user interaction, the "Retry all" button should have low affordance.

BEFORE

AFTER

Frequent failed GT uploads and ambiguous data quantities were difficult to identify.

Data uploads are critical to project progress.

Interviewed internal labelers to prioritize and address the most significant UX issues first.

If the failure reason is long, it should be downloadable.

For optimal user interaction, the "Retry all" button should have low affordance.

Solution 2

Visualizing the AI Data Curation Process

Embedding visualizations offer deeper insights into model performance and data quality beyond traditional metrics. This is why our team focused on developing these visualizations.

Drag the cursor to select and display the thumbnail images.

Solution 3

Synthetic Data Integration for Superior Curation

Leveraged synthetic data to improve data quality and bolster the curation process.

Key Features

Meta Filter Options

Designed users to adjust the meta values (e.g., time of day) for the generated images, providing control over the specific characteristics of the synthetic data.

Advanced Settings

Offers the ability to choose different models (e.g., Stable Diffusion V1.0) and customize the number of images generated, giving users flexibility in the data creation process.

User-Friendly Interface

The design ensures that users can navigate between original and generated images effortlessly, making it simple to mange and curate large datasets.

Solution 4

Model Deployment Workflow Creation

I worked with my team to develop the Model Deployment Workflow, a crucial part of the AI model development process. This workflow was built from scratch to guide users through deploying AI models efficiently and intuitively.

Comprehensive Model Setup: From Dataset Configuration to Augmentation

Augmentation Step in Model Workflow

User can select various augmentation options, such as flipping, rotation, adjustments, to enhance the variability in the dataset. This ensure that the model is trained on a wider range of data, improving its accuracy and generalization.

The augmentation settings are seamlessly integrated into the model creation workflow, allowing users to easily apply these transformations without disrupting the flow.

Team Decision

My team and I decided to incorporate this augmentation step to generate richer datasets, ensuring the model is trained on a larger and more diverse set of data. This enhancement is crucial for producing models that perform well in varied real-world scenarios.

Impact

Incorporating augmentation at this stage enhances the dataset by introducing controlled variability, leading to a more resilient and well-rounded AI model. this step is vital for creating high-performance models capable of handing diverse real-world scenarios.

Solution 5

Model Training: Post-Training Analysis and Evaluation

With just three months to deliver an MVP for CES, I collaborated closely with the Labs and AI engineering teams to design the first phase of model deployment. Despite having no prior experience in model training or data curation, I quickly learned the essentials and focused on creating a clear, usable interface for post-training analysis.

Graph Analysis

User can access detailed graphs that depict the model’s performance over the training period.

Evaluation Section

The evaluation section provides a visual comparison between the Ground Truth and the Expected Checkpoint.

Takeaways

Takeaway 1

Focus on High-Impact MVPs Under Tight Deadlines

"Never underestimate the power of a tight deadline."

With only three months until CES, I focused ruthlessly on prioritizing core MVP features that would deliver the most business value. I quickly adapted to complex AI/ML domain by collaborating daily with AI engineers to successfully ship a functional prototype on time.

Study with AI Labs team for Meta Filter Feature

2. Treat Internal Teams as Users

"Internal Users are Users, Too."

I learned that great UX doesn't stop at the customer. By designing clearer screen flows and documentation for our developers and QA team, we ensured stronger alignment and smoother handoff, which directly boosted internal efficiency and team morale.

Visualizing complex screen flows to align Development and QA teams.

3. Innovation Value Award: Design Operations

"I realized that the 'product' isn't just what we ship to customers, it’s also the system our team uses to build it."

I spearheaded the implementation of new standards to eliminate cross-functional friction and improve our design practice. This included creating a Figma File Organization & Project Documentation Policy for the design team and establishing new developer handoff protocols. This systemic contribution earned the Design Operations Achievement Award 2023, providing that designing for internal alignment is key to scaling a product organization.