Modernizing Photobucket's React Native App – and the Team Behind It

Upskilling the Photobucket team along the way

Embedded collaboration also meant making sure Photobucket's engineering team could benefit from our expertise in the React Native ecosystem and pick up more good practices along the way.

We ran a continuous knowledge-sharing track alongside every phase of the project: regular code reviews with detailed written commentary on the patterns we noticed and how to improve them, observations shared in Slack as we worked, and a dedicated training session on React Native Reanimated – a powerful animation library that can hurt performance when misused.

We made sure the team understood not just what we were building, but why we were building it that way, so they could carry that knowledge forward independently.

Turning a FaceFinder proof-of-concept into a production feature

Photobucket's FaceFinder lets users tag a person in one photo and automatically surface every other photo of that person across their Groups. The client had a fast, working proof-of-concept when they came to us; what it needed was to be turned into a production-ready feature.

To make the right architectural decisions for production, we started with extensive benchmarking before touching production code. We compared multiple inference runtimes, face-detection approaches, and recognition models across accuracy, false-positive rates, and latency, giving the client team clear, data-driven evidence for every major decision. We evaluated both detector models and native platform solutions for locating faces in images, plus multiple recognition models for converting detected faces into the vector embeddings used to match people across their Groups.

Beyond model selection, we improved the post-processing pipeline around face detection, optimized native inference performance on both iOS and Android, and wrote the native code responsible for fully local inference using ONNX Runtime.

Once the benchmarks made the case, we settled on a recognition model architecture well-suited to production – one that performs reliably on real-world photos, handling variations in lighting, angle, and image quality. FaceFinder now runs entirely on-device – fast and accurate.

A new design system and a carousel built with structured AI engineering

Partway through our collaboration, Photobucket rolled out a new design system that changed the look and feel of the app entirely. We helped build it – creating the new navigation from scratch, new component primitives, and an integration with React Native Header Motion, a wrapper around Reanimated and Gesture Handler that makes animated header behaviors straightforward to implement.

The most technically interesting part of this phase was the photo carousel: a full-screen, gesture-driven viewer with pinch-to-zoom, focal-point awareness, immersive mode, high-resolution image loading on zoom, overlay support, and animated headers and footers. It required us to get both the design and the business logic right while handling a number of subtle edge cases.

We approached it with a structured agentic engineering workflow that kept quality high without sacrificing speed. Even with a detailed, well-documented vision for the carousel, Claude Code's initial output didn't reflect how we wanted to approach the architecture – and without a shared spec, there would be no reliable way to course-correct. So we invested more heavily in the spec upfront.