A published send ceiling per domain, enforced by the product.

¶01Viktor, our CEO, had eleven sending domains and a sequence that worked for nineteen days. By day twenty-four, four were landing in spam on Microsoft 365. Nothing in his tooling had said no: it offered more mailboxes and let him raise volume the week two people asked to be removed.

¶02Plimsoll runs sending domains and mailboxes under a published daily ceiling, and reads the replies to decide how fast each domain climbs toward it. The ramp takes 21 days, and no support ticket or contract raises the maximum. We own the ramp model, the reply classifier and the labeled corpus the classifier learns from.

¶03Plimsoll's reply stream is where header rules run out: bounces and auto-responders carry structure, but roughly one reply in six is free text. A model classifies and scores those, an unsubscribe that also sounds interested or a forwarded complaint, and quotes the sentence it decided on, retrieving against an index where every reply is embedded. Without it, each would hold its domain's ramp until a person read it.

¶04Any reply under Plimsoll's confidence floor stops its domain's ramp and waits for a named reviewer. Holding a domain flat for a day costs little; misreading a complaint costs the domain. Each reviewed reply becomes a labeled pair, the quoted sentence beside the reviewer's decision, detached from customer and recipient identifiers. The pairs outlive the 180-day deletion of reply bodies; our classifier is fine-tuned on them.

¶05Plimsoll's GPU hours pile up at onboarding and in a two-hour daily send window. Onboarding embeds months of a customer's reply history in one batch; in the window every customer's replies are embedded and scored within two hours, and the machines idle either side. Mailboxes run on cloud infrastructure in Oregon, the reply store and corpus in object storage there, and the embedding pass on GPU instances switched off outside the window.

¶06From Q1 2027 Plimsoll's own classifier, an open-weight 3B model trained on the reviewed pairs with Azure Machine Learning, reads the free-text replies on our GPU instances in Oregon. Until then Azure OpenAI in West US 3 classifies them, under zero data retention. Our infrastructure bill in Oregon also grows with every domain: more mailboxes, a larger reply store, more instance hours. Both regions are in the United States, a condition buyers set before handing over a domain.

¶07The three of us incorporated Plimsoll in Delaware in January 2021 and have paid our own way from Austin since. Nine customers pay $90 a month for each domain at its ceiling, flat, never by volume. The onboarding backfill, free to the customer, is our heaviest compute, and it must stay affordable as domains per customer climb.