Projects
WORK
Selected projects spanning AI infrastructure, design systems, vehicle transport, and roadside assistance.
AI STORAGE CONSOLE
MinIO AIStor is a high-performance object store purpose-built for AI and analytics workloads at any scale. The AIStor Console provides a browser-based interface for managing the AIStor platform — bringing the full power of the command line into an accessible, visual product.
Initial Product
Led the end-to-end creation of the Figma design system from the ground up, setting the products foundation. Working across design and engineering, I drove both the overall design direction and the day-to-day execution — collaborating closely with a cross-functional team to move from concept to shipped product.
To ensure we built the right thing, I interviewed engineers and stakeholders across technical levels — letting that research shape both the feature set and how the product communicated to users of varying expertise. I also ran Figma learning sessions to level up the team and keep everyone moving fast.
shadcn/ui Redesign
The redesign wasn't just a visual refresh — it was an opportunity to rethink how design and engineering worked together. I built an entirely new design-to-development workflow that kept Figma and code in sync, and used Claude to rapidly prototype explorations across color direction, navigation, layout, and component patterns. What used to take days of iteration could now be evaluated in hours.
Claude became a core part of the process in a deeper way too. I developed design rules that let explorations start directly in code while preserving consistency across the product — then led engineers in using that same workflow to design feature sections themselves, pairing their domain knowledge with a shared design language. To close the loop, I built a code-to-Figma process that automatically updates designs from production code, eliminating drift between what's shipped and what lives in the design file.
Work Examples
WEBSITE DESIGN SYSTEM
The MinIO website had grown organically over time, resulting in a fragmented visual language and an increasingly difficult-to-maintain codebase. I led the initiative to build a comprehensive website design system — establishing the foundation for a full site redesign while solving for long-term consistency and scalability.
A core goal of the system was to make content creation accessible beyond the design team. I designed the component library so that marketers, writers, and other non-designers could assemble pages confidently, without needing to reinvent layouts or second-guess visual decisions. This significantly reduced the bottleneck of routing every new page through design.
I drove both the design direction and the team's execution — establishing component standards, leading education efforts, and ensuring the system was adopted effectively across disciplines. The result was a shared visual language that made the redesign faster to build and easier to maintain going forward.
Work Examples
ACCIDENT PHOTO APP
Exploring the Problem
Not enough tow operators were using the initial web form. When they did, many did not take the right kind of photos before towing the vehicle. To understand the reasons, we looked over submitted photos and recorded why each was not acceptable — too close, missing a portion of the vehicle, etc. We examined detailed job data to understand situational factors. Lastly, we interviewed tow company owners, dispatchers, and tow operators to learn more about their individual experiences.
Opportunities & Considerations
Existing process confusion
Many tow operators already take photos to protect their company from lawsuits. When asked to upload photos, many uploaded images they already took — which often didn't meet HONK's requirements.
No perceived value
Uploading photos was an extra step with no apparent benefit to the tow company or operator — making it easy to skip.
Photo taken after tow
The value of this product comes from documenting the vehicle's state before it's towed. The experience was doing nothing to enforce this timing.
Ineffective instructions
Whether ignored, glanced at, or misinterpreted — users consistently weren't following the instructions provided.
Safety
Roadside assistance is already dangerous. Taking the requested photo often required stepping into a hazardous area — a highway, steep hill, or active intersection.
Capturing the wrong person
Tow companies were notified via text, which often reached a dispatcher rather than the operator physically at the scene.
Camera orientation
Some operators held their phone vertically, making it nearly impossible to capture the full vehicle in a single frame.
Tow operators are not always contacted about HONK's request to take a damage photo when assigned an accident tow job, and when they are, they are not taking damage photos that meet HONK's criteria. Tow operators need to be informed of the photo request on every accident tow job, encouraged to do it, and guided through the photo experience so HONK has an acceptable damage photo for total loss assessment.
Testing & Iteration Strategy
Once we had a usable first version of the camera experience, we used split testing to validate each iteration. The scope of each iteration was focused on changes that could be validated with confidence — usually small, resulting in 1–3 day cycles. Validation was accomplished using a funnel representing the life cycle from when an Accident Tow Job was first created to when an acceptable photo was taken. We also used screen recordings and user interviews to understand the "why" behind any funnel drop-off.
Getting Better Pictures
We prioritized photo quality first because we believed it would require the most technical research, and wanted to understand the options and limitations early — as they would influence every subsequent decision.
We first explored a guided video walk-around, extracting stills for damage assessment. Browser limitations combined with poor mobile data speeds forced us to abandon this approach. We then shifted to a guided single-photo experience, hoping to create a more reliable path to an acceptable image.
Building an Intelligent Web App Camera
Live object detection
The idea was to use a camera that could detect the vehicle in real-time and provide live feedback as the operator framed the shot. Two critical constraints were device performance and internet data speeds. After many iterations and extensive internal testing, we determined that a viable product could not be built for enough of our users across the varied locations where the app would be used.
Object detection after capture
After ruling out live detection, we redesigned the experience so users submit the photo they took and receive automatic feedback on whether it was acceptable. When rejected, the user was reminded of the guidelines and prompted to try again.
To prove the design concept before engineering the backend, we sent all submitted photos to a Slack channel where team members manually approved or denied them in real time. This was released to real users to validate before building automation.
Camera Orientation
Without live feedback during capture, it became even more important to guide the operator upfront. This screen prompted users to hold their phone horizontally before activating the camera — improving the likelihood of capturing the full vehicle in a single frame and increasing acceptable first-attempt submissions.
Increasing Engagement
After improving photo quality, our next challenge was increasing how often tow operators used the app during accident jobs. The existing text-with-link integration was ineffective — it was sometimes sent to the wrong person, often ignored, and created a disjointed experience. A substantial number of operators also weren't using the HONK Partner app at all (30% of Accident Tow Jobs), so improvements needed to reach both groups.
40→75% HONK Partner App Integration
We displayed the photo assignment and a direct link in the same section where operators viewed job details — leveraging existing engagement with the job view. We also added a push notification when operators arrived at the scene as a timely reminder. This integration had the largest single impact on engagement, raising it from 40% to 75%.
+3–5% Improved Handoff
When the tow operator wasn't the one using the HONK Partner app, the dispatcher received the notification instead. We added a simple way for dispatchers to forward the photo web app link via text to the right person at the scene. After this change, we saw consistent usage of the forwarding feature and a 3–5% lift in overall engagement.
Value for the Tow Company
Proof of existing damage protects tow companies from liability claims. We added photos and video taken through the web app to the operator's job history page, giving them quick access to documentation if they were ever blamed for vehicle damage. We didn't see immediate engagement improvements, but received positive responses from tow company owners.
Increasing "On Scene" Photos
After the Partner app integration, 4% of photos were still being taken after the vehicle was towed. Operators cited safety (not wanting to step into a busy road) and situational factors (police directing immediate removal) as the main reasons. Not wanting to suggest anything unsafe or that would strain HONK's relationship with operators, we decided to leave this unchanged.
Takeaways
More value for tow companies
Value to the tow company and operator was one of the least explored opportunities, with substantial potential to increase engagement — and to strengthen HONK's relationship with operators in ways that extend beyond photo app usage.
Browser-based cameras are problematic
The browser camera offered a controlled experience — visible framing, aspect ratio enforcement, and integration. But its reliability and device permissions dependency created friction that native cameras don't have, converting fewer visits to photos taken.
Post-tow photos still have value
The project focused on pre-tow photos due to their higher financial value. However, photos taken after towing would still assist with claims processing and could be added to the experience to improve the overall product offering.
DROP-OFF VERIFICATION
Problem
Carvana's vending machine car delivery is a signature experience — but the vehicle drop-off verification process was entirely manual, error-prone, and creating significant operational bottlenecks. Inspectors were using paper forms with no standardization.
Causes
Tow operator indicates vehicle is delivered prematurely.
Tow operator marks the wrong car as delivered.
Carvana agents are not present during drop-off to mark where the vehicle was unloaded.
Solution
Verify the vehicle is at the correct Carvana lot (via GPS and vehicle VIN), and identify where it was left in the lot with GPS coordinates and a photo for visual assistance.