Monitoring interpretability in production

Interpretability gets framed as an offline research activity. I think that’s a mistake. The same feature activations you study in a notebook can run as a live signal in production.

In GlassBox, every inference emits its probe scores and top feature activations to Arize Phoenix, with Sentry catching the anomalies. When an overconfidence probe spikes, that’s an event — not a metric you discover three weeks later.

The infra is boring on purpose: a FastAPI service, a RunPod GPU worker, and a thin telemetry layer. The interesting part is what you choose to watch.

(Placeholder post — replace with your own writing.)

← all transmissions