Writing and voicing a session for one person, with the bill under control
Every session is written and spoken for one person and one moment. Nothing is pre-recorded, which means every single play costs money to produce.
A product that generates fresh audio per request has an unusual failure mode: it works beautifully and then the model bill arrives. Cost is not an afterthought here, it is a design constraint that sits next to the writing.
The pipeline takes what the person says they are carrying, drafts a script to a structure that holds the pacing and the silences, then voices it. Each stage has a budget. If a stage runs long or retries, the run is capped rather than allowed to spiral, and the queue smooths bursts so a spike in demand does not become a spike in spend.
Model choice is per stage and gets revisited. The model that writes the script is not the one that voices it, and neither is fixed.
What runs now: request to script to voice to delivery, with a per-run cost ceiling, a retry policy that fails closed instead of open, a queue that absorbs bursts, and per-stage model selection that we change when something better or cheaper appears.
What it taught us: cost control is a feature of the architecture, not a report you read afterwards. Every client automation we build inherits the same ceilings.