CrumbleUX

Most product feedback arrives too late. A recording is collected, a researcher watches it later, and the useful observation is separated from the moment that caused it. I built CrumbleUX to make that loop immediate: it watches a product session, listens to the person using it, and responds while the context is still visible.

The product brings six UX perspectives into one session. Each one looks for a different failure mode, from unclear hierarchy and accessibility problems to hesitation, trust and conversion friction. The aim is not to imitate a room full of designers. It is to make critique specific enough that a team can decide what to change next.

I built and operate the system end to end. The application is TypeScript and Next.js, the inference layer is Python, and the product handles live screen and audio streams, accounts, teams, billing and the less visible work required to keep user sessions private and reliable.

Streaming is not the hardest part. A model can describe almost anything on a screen; useful critique requires knowing what matters, avoiding repetitive observations and grounding each comment in the interaction that prompted it. Much of the work has been turning plausible commentary into advice a product team can act on.

CrumbleUX is an active product, not a claim that automated critique replaces research with real users. It is most useful earlier in the loop, when a team needs another set of eyes quickly and can still change the product cheaply.