CASE STUDY : Hyundai
Hyundai
SODAS: SAFETY OFFICE DATA ANALYTICS SYSTEM
Leveraging GenAI/ML and Big Data to recall parts, systems and vehicles more efficiently.
My role
Product Design Principal
What I Did
01
Project Scoping
02
Requirement Gathering
03
Stakeholder Collaboration
04
Engineering Whisperer
05
Workflow Mapping
06
UX
07
UI
08
Design Sprint Facilitator
09
User Interviews
10
Team Consensus Builder
“Greatly appreciate your efforts throughout the project. Your work has been appreciated a lot by Hyundai – they loved the entire design experience that you led.”
Where we were headed
“This was a project at Hyundai where we were designing for 40 Data Analysts in Irvine, CA who needed to recall parts, systems and vehicles more efficiently. I was the principal product designer on a cross-functional team with an 80 person team-Big Data, Server, front end, UX, UI, leaders from 5 companies, engineers, etc. The business goal was to leverage data (+50 million records) to recall vehicles faster and make the US government happy, and my role spanned from research to execution.”
Hyundai Motor Group
-
Hyundai
-
Kia
-
AutoEver IT solutions provider
-
Mobis Parts Company-
How it got started.
Hyundai Motor Group put out an RFP for a new complex AI system to use the massive amounts of data available to detect trends and recall parts & systems from Mobis, vehicles from Hyundai and Kia and house it all at AutoEver on one of their internal servers.
InfoSys/WongDoody won the bid.
Infosys lacked the resources inhouse, so they called me.
I am a car guy. And I love gaining consensus and seeing a vision come to life.
When I was hired, Hyundai Irvine was not getting their vision across to the offshore InfoSys team. The SOW had yet to be signed due to conflicting visions as to what this massive system would do.
Add to this- in Nov 2021 we were still in the throws of Covid and entire teams in India were calling in sick. Plus, there was a chip shortage and there was no way to stand up a physical server, and the decision was made to migrate the project to AWS.
PLUS
the Data Analysts were not happy with the pace or direction of the SODAS project. They were worried that it was just another tool they would be forced to use that did not make their job easier or better.
SO
I was hired as a Local Automotive Subject matter expert, Product Principal and as cultural liaison.
Kickoff
Starting off in a fairly chaotic manner, the project was in need of a NorthStar vision. The Program Manager at Hyundai, who had spent 18 months planning and setting the whole thing up, was retiring. I saw that no one but her had a good grasp on the scope and complexity of this endeavor. So, I asked her to sit with me and share all of the details with me. Within a few days, I became the project authority so she could retire and know her vision would be realized.
ALSO
I read the original RFP from Hyundai, the proposal from InfoSys that won the bid and looked at the staged rollout of the technology. Then I spoke with each team lead and all the stakeholders. I discovered a lot of variation in the vision and what the technology would do.
From there I came up with a plan.
Using what I learned from the Program Manager, I became the touchstone.
By the time the PM retired, we had a clearer vision and game plan. I started preparing visual artifacts for every meeting. And asking lots of questions. I worked out flows and processes, then team workflows and integration with internal tools.
During discovery, I saw that the various data source were all on different platforms and the Data Analysts had to retype or copy and paste into Excel Spreadsheets. Then I discovered a process I called the, Yo Eric.
Yo Eric
During a discovery session with a manager of one of the Data Teams I was asking questions about the data gathering process.
What do you do next?
"I walk across the room and login to that computer and see what information is there. It is not connected to my workstation. Then I copy and paste into my excel spreadsheet."
OK what do you do then?
"I yell, Yo, Eric! Check out the spreadsheet I posted on the server and tell me if we have a case"
We could do better.
LOL
Where we were headed
“This was a project at Hyundai where we were designing for 40 Data Analysts in Irvine, CA who needed to recall parts, systems and vehicles more efficiently. I was the principal product designer on a cross-functional team with an 80 person team-Big Data, Server, front end, UX, UI, leaders from 5 companies, engineers, etc. The business goal was to leverage data (+50 million records) to recall vehicles faster and make the US government happy, and my role spanned from research to execution.”
After the PM retired, I was able to keep her vision alive while enrolling all the other teams in the NorthStar vision and the roadmap. I kept producing artifacts for every meeting. Morning with the Irvine and Seattle teams and every night with the India teams. I managed a UX design team in India. They would socialize through the InfoSys teams what I had created and gotten approval for during the day. I would record all the meetings where I showed the artifacts (process flows, workflow diagrams, wireframes and other visual representations of the current state of the SODAS project).
At some point, the AutoEver leads would say, "Ok Bryan, show us what you got today"
And I would facilitate the meeting and get buy-in, opinions and consensus. I MADE THEM LOOK GOOD. AND Hyundai, Irvine was starting to like the direction we were going in.
I love my job. :)
SUCCESS
I was able to gain consensus
with all 80 team members
on the NorthStar Vision and
roadmap for a successful product.
Objective
Create a comprehensive alert system based in existing "HAZARD" assignment and taxonomy that would allow Hyundai to recall vehicles, systems and parts more efficiently with AI and big data.
HMA has resolved the inquiry by National Highway Traffic Safety Administration (“NHTSA”) regarding the timeliness and scope of recalls in compliance with the consent order.
Hyundai is committing to implement new IT systems to better analyze safety data and identify potential safety issues. Development of systems directly benefiting and supporting the safety office to be best in class U.S. safety office. As part of this commitment HMA Safety Office needs to integrate machine learning and predictive analytics into existing processes to identify and investigate potential defect trends.
GOAL
Become more proactive in vehicle recalls with AI.
CONCEPT
1- Machine learning would injest 14 million records from service bays and warranty claims in Phase 1, and would detect trends based on KNOWN HAZARDS and throw alerts.
2- The analysts would:
-
see alert in feed
-
do extensive research across all data sources
-
compile data in a container
-
make recommendations to Case Management Tool
3- Results would get peer reviewed
4- A case would be started in the CMT.
5- Parts, systems and/or vehicles would be recalled.
What Hyundai thought they were buying
They expected the system to sift through 50 million records and identify action areas, pinpoint trends of failure, and determine which parts, systems, and vehicles needed to be recalled.
But that assumption was wrong.
In reality, you first have to instruct the GenAI on what to look for, define what counts as success or failure in the context of HAZARD alignment, and then synthesize large amounts of information before the model can be trained to recognize patterns and report meaningful findings.
Where we ended up:
Alert Feed plus
Research Toolset
When it was determined that ChatGPT failed to find any trends, everyone froze. This is where I dug in.
Any tool has to work with no data, too.
I created a toolset that worked with zero alerts from the system as well as thousands of alerts.
What I learned through DISCOVERY, RESEARCH and Interviews.
1- Creating a portal with all 15 data sources in one place was a huge lift and a critical to the Data Analysts adoption of the tool/toolset.
2. Simplifying and clarifying what a HAZARD is
AND defining what an alert was was a challenge and needed to be done at the beginning for a successful toolset.
3. The Data Analysts were discouraged and had low confidence in the value of yet another tool, especially without their input.
4. The toolset had to parallel their current WORKFLOW and we could get great leverage with simple improvements and modernization.
5. After the design sprint with the Data Analysts, they had a renewed faith on the success of the new SODAS system.
Phase 1
Data
Sources
With over 40 data sources, Hyundai chose 2 large data sources for Phase 1.
Service bay records and warranty claims.
14 million Service Bay records
The front line data came from the Hyundai dealership
service bays.
Chat GPT
ChatGPT (2.5) in 2022 was very rudimentary and had no consistent interface.
Especially with very large
record sets.
We learned a lot in the first phase.
• How Safety Analysts gathered research
• How they kept it in one place
• How they handed off to the Case Management Tool (CMT).
Discovery
ChatGPT had a few issues with service bay records
When Hyundai mechanics and technicians took notes and/or wrote up a service bay record, they didn't always use proper nomenclature.
Example- "I told her to fire up the engine, but she said it was broke."
Is this a Engine Bay Fire hazard?
Toolset- Advanced search with saved search, documents attachment and
pre-loaded quick search.
Search
Quick Search
I created a new tool.
During the Design Sprint with the Data Analysts, I discovered that the research phase took hours or days per case, so I created a few tools that helped them more quickly search. This Quick Search tool allowed anyone, with the click of a button, to go from an alert to a search with all of the Vehicle/Year filters selected and the correct HAZARDS selected.
Advanced Search as Research Collection Toolset
Search contained all documents (warranty claims, service bay records, etc.) plus any existing alerts. Agents could select any items found in search and "collect" them into a bucket.
Later renamed to Case Builder.
Integration through API to existing
Case Management Tool was key to
SODAS product success.
Case Management Tool Integration
Case Management Tool API Integration
Although it was not in the original SOW, integration to existing CMT through the API was mandatory for data analysts to do their jobs. My design workshop illuminated their true workflow.
I designed a way to create a CMT case from any alert.
Alerts could show up in the thousands or not at all.
The feed is a way to manage, sort, filter and even create an alert. Alerts and the feed became part of the research workflow.
Alert Feed
Alerts-
Containers for systems,
parts, research, documents
and notes
Using the alert as the container, allowed the Data Analysts to gather research as they moved through their days and their workfflows.
Train the model. Aligning service bay records, warranty claims and other documents with the HAZARD TAXONOMY was crucial for the success of the trend finding alert system.
Train the Model RLHF
Training the MODEL
We were lucky.
We had the experts as users.
The Data Analysts were driven by the need to be the authorities and being good at their jobs.
They were perfect for training the model. HAZARD Classification was the crux.
First - we used their self-made alerts to train the model.
Then - we had them to align claims and records to specific known hazards. So we put them to work to train the model by assigning claims and service bay records to specific HAZARDS.
They were exceptionally competitive.
I put that to work too.
I gamified the Train the Model section. They got so excited!
Where we ended up.
My role was complete after 6 months.
We had a NorthStar vision, a working prototype of the entire system, a design system with Hyundai branding, a model that was learning, a project roadmap, a signed SOW
and a happy client.
"Thanks for making it so easy to use."
Data Analyst during design sprint results presentatation of my designs
What I learned
-
AI has to fit into a workflow. Mapping the actual workflow was the golden ticket to a successfu toolset.
-
The toolset has to work with no data as well as LOTS of data (14 million records).
-
Enterprise solutions have a lot of inputs but not always an obvious solution. But the solution is findable and achievable.
-
Without doing the UX process completely, you don't know what the toolset needs to do.
-
I naturally fill leadership voids and I am fun to work with.
-
I can see the vision and use artifacts and facilitation to get everyone on the same page moving forward.
-
AI is challenging.