About Understudy

Understudy is built for whoever owns the demo library at a B2B SaaS company of sixty to four hundred people, and gets asked by sales for one-off versions: the integration screen with the prospect's own systems named on it. The category prices the library rather than the use of it. Twenty demos at $19 a month, fifty at $49, past that an Enterprise plan with no price. So the count binds in week six of the quarter, somebody deletes the two demos with the fewest views, and their engagement history goes with them.

We are one endpoint. You record a flow once and post a recording id with a context object: the prospect's company, their industry, the systems they run, a handful of figures. Back comes a demo URL with its own analytics and its own expiry. Nothing is generated. Every screen came from your recording, with certain spans replaced. The name in the header becomes theirs, the chart figures become the ones you supplied. The rest is byte for byte what you approved, which is why an account executive can send it without watching it first.

The classifier does the job rules cannot finish: which spans are substitutable, and which are chrome that must never move. Exact matching catches the seeded account name where it stands as itself, about two thirds of the time. It misses the possessive in body copy, the name inside an alt attribute, a figure that is part of a computed total. Below 0.95 confidence nothing is substituted and the span is flagged, because the failure this product cannot survive is a demo showing a prospect somebody else's name. A named person rules on the queue, and that approval trains the classifier for that account alone. Only the layer beneath crosses accounts, meaning whether a region is a header cell, a chart axis or body copy, and it carries no string of yours.

Compute is bursty on a schedule nobody controls. One demo is cheap; ingest is not, because a fifteen minute capture is tens of thousands of candidate spans classified in one pass. The other burst is a campaign: three thousand demos in an afternoon, then nothing for a week. Both land on AWS in Ireland, eu-west-1: ingest, rendering and the classifier run on EC2 there, with on-demand GPU instances serving the models we operate, and recordings and generated demos sit in S3. That bill grows with every demo generated. Training does not need to sit next to a prospect's click, so it runs in Ireland as well, on Azure Machine Learning in Microsoft's North Europe region: the substitution classifier learns from each account's reviewer approvals, and the interface taxonomy from corrections shared across accounts. Through 2026 the classifier runs on an open-weight vision backbone, and a hosted endpoint sees only offline evaluation against captures we made ourselves; from the first quarter of 2027 that evaluation may run on Azure OpenAI in North Europe, with nothing retained. Through 2027 the backbone is fine-tuned on the interface taxonomy in Azure Machine Learning and served back on EC2, to bring ingest for a fifteen minute recording under sixty seconds. That is the number to hold us to.