Understudy
Sales keeps asking for one-off versions naming the prospect's own systems. The category prices the library: twenty demos at $19 a month, fifty at $49. The cap binds in week six of the quarter, and somebody deletes the two least-viewed demos with their engagement history. Understudy's three founders started Understudy S.A.S. in Bogotá, Colombia, in May 2023.
You post a recording id with a context object, the prospect's company, industry, systems and figures, and get back a demo URL with its own analytics and expiry. Every screen comes from your recording; only the header name and chart figures change, and the rest is byte for byte what you approved.
At ingest every frame is segmented, and a vision model trained on screens, not text, gives each candidate span a type, a substitution rule and a confidence. Exact matching finds the seeded account name about two thirds of the time; the rest, the possessive in body copy, the alt attribute, the figure inside a computed total, is the classifier's job.
Each approval or correction becomes a training pair, the span in its screen context against the confirmed rule; a fifteen minute capture sends forty to two hundred spans to the queue. The pairs train that account's classifier alone. What pools across accounts is the region type beneath them, header cell, chart axis label or body copy: an interface taxonomy Understudy owns.
A fifteen minute capture is tens of thousands of candidate spans classified in one pass; a campaign is three thousand demos in an afternoon, then nothing for a week. So the classifier runs on on-demand cloud GPU instances in Dublin, Ireland. A recording holds whatever tenant it was captured against, so none is sent to a hosted model endpoint.
The endpoint and the review queue are running, and ingest accepts captures under twenty minutes until December. Through 2027 the open-weight vision backbone is fine-tuned on the interface taxonomy in Azure Machine Learning's North Europe region, also Ireland, and served from Dublin, where recordings and demos sit in object storage.
Andrés Galvis
CEO
Ran demand generation at two B2B companies and spent both of them arguing with a demo library about which fifty demos were allowed to exist that quarter.
Kate Deshko
Founder
Carried a quota for six years, built one-off demos by hand in a screen recorder and sent them, and stopped after counting the hours against the deals they moved.
Sebastián Cardona
CTO
Built the capture and replay layer at a session-recording company, and knew that a recorded screen is structured enough to edit safely and unstructured enough that everybody assumes it is not.