See the finished product first.
A worked sample of what we write, and the evidence pack your attorney receives for every published piece.
Sample article · not a published piece
The engineer who cut a model's inference bill by 60 percent
A worked example of what we write: about a person, built on figures somebody else could check, with the claim and its evidence in the same paragraph.
Most teams respond to a rising inference bill by buying more hardware. When the cost of serving a mid-sized language model at a 40-person company doubled in a quarter, its lead engineer did the opposite and went looking for the waste.
The answer was in scheduling. Requests were being batched on arrival rather than by shape, so short prompts waited behind long ones and the accelerator sat idle between them. Rewriting the batching layer to group by sequence length lifted utilisation from 38 percent to 71 percent, measured over a fortnight of production traffic.
“The hardware was never the constraint,” they said. “We were just feeding it badly.”
The change cut monthly serving costs by roughly 60 percent on the same traffic. The approach has since been published as an internal engineering note and adopted by two other teams at the company.
Read this bit. This is an illustrative sample written for this page, not a published article, and the person in it is invented. It shows the shape, the register and the standard of evidence we write to.
The evidence pack your attorney receives
One of these per published article, delivered as a set rather than as links in an email.
- Full-page capture
A dated PDF of the article as published, including the masthead.
- The live URL
Plus an archive snapshot taken on the publication date.
- Headline, date, author
The three things the regulation names, set out on one line.
- Author page or biography
Where the publication provides one.
- Publication background
What the outlet is, who it is for, and how long it has published.
- Audience figures
Supplier-reported monthly visits, labelled as such rather than as readership.