building

Models get the attention. Data gets the hours. At Pixa I worked the layer underneath: scraping, ingestion and preprocessing systems that had to keep running at volume, not just run once in a notebook.

Multi-GPU encoding pipelines, because the bottleneck was never the model. Most of the work was finding where the real constraint had moved to this week — disk, network, a single-threaded decode step nobody had looked at — and moving it somewhere else.

Retrieval-based voice conversion trained and deployed on top of that infrastructure, with pipelines scalable enough to handle speaker adaptation as a routine operation rather than an experiment.

And dashboards, throughout. A pipeline you cannot see is not a pipeline you have — it is a pipeline you are hoping about. Tracking data quality turned out to matter more than tracking uptime.

I have been assembling things since I was a kid — Discord bots, websites, npm packages, whatever the current obsession was. Python, JavaScript, some Rust. It is the same instinct pointed at larger machines.

Why is data infrastructure still the least written-about half of machine learning?