This is where things got interesting.

I spent a year at CR3ATE.AI. For the first six months I was a developer intern, and for the last six I was a full-time junior developer. We were a two-person team, and we built the stack, got it all working, and shipped it to customers ourselves.

CR3ATE.AI works on computer vision, AR/VR and real-time ML. There was no gentle ramp-up. I got thrown straight into the deep end, which was exactly what I needed.

Building and shipping the product

Most of my time went into the product itself. I co-built the production iOS app and the customer-facing web frontend and backend for cr3ate.ai, and connected them to the ML pipeline through our APIs and AWS.

On the backend, I built NestJS APIs, containerised with Docker and backed by MongoDB. I wrote AWS Lambda functions for file processing and handled uploads to S3, which fed the processing pipeline. Firebase handled auth. A lot of the work was getting JSON to parse cleanly from one service to the next while the product was shipping.

The job was close to forward-deployed engineering. I was customer-facing, I built the stack customers used, and I got it all working for them. On a two-person team you can't stay in one layer of the stack.

It was also my first time working on something live and customer-facing, and it changed while people were using it.

Oh, and the app I co-built got through review and went live on Apple's App Store. If you know how harrowing Apple's approval process is, you'll know that's an achievement in itself.

The ML underneath

The product sits on ML, and I worked on that too. I developed CoreML models for iOS, specifically real-time blur detection in camera apps. I evaluated and optimised architectures like MobileNetV3 and EfficientNet in TensorFlow, benchmarked them against each other, and worked out which ones held up on a phone. On a phone, latency and model size decide whether a model is usable. Most ML work never forces you to care about either.

I also integrated Hive APIs to support 3D Gaussian splat reconstructions for AR, VR and mixed reality workflows. Gaussian splatting is one of the most exciting things happening in 3D scene reconstruction right now. Getting that pipeline working end to end is still one of the most satisfying technical things I've done.

The ML was fun, but shipping production software on top of it taught me different lessons. Applied ML and full-stack infrastructure overlap in messy ways, and CR3ATE.AI is where I got comfortable in that mess.

Where it went

Before I signed off we'd started on agentic workflows. These are automations that run from a single prompt, plus self-healing loops for when a reconstruction doesn't come out just right (artists are finicky, what can I say).

PXO (Pixomondo), a Sony subsidiary, demoed our stack at CES 2025. CR3ATE has also worked with Netflix, Amazon, a lot of government contractors and large VFX studios.

See what they do at cr3ate.ai. My certificate (my badge of honor) covers the internship half. The full-time junior developer role came after it.