Blog
What we shipped.
Product updates, engineering writeups, and the occasional reflection on how we got here.
All posts
Latest
We cut our inference bill 60% by changing one string
2026-05-26
Atlas-1 is now Newmen's optimization-layer default. Drop in the model name, and Atlas serves the cheapest path that passes your eval gates per call. If quality drops, the call isn't billed.
3 min read
Introducing Atlas
2026-05-16
One env var. The models you already use. Atlas is the optimization layer that serves every call the cheapest way that holds quality — and proves it on your own traffic.
3 min read
Why we built Newmen
2026-05-12
A founding note on model drift, the annotation gap, and the three things we decided you should never have to do yourself.
4 min read
Prompt engineering is a local maximum
2026-04-28
Why iterating on prompts solves the wrong problem, and what the production-safe alternative looks like.
4 min read
Ship gates caught a regression we didn't know we'd introduced
2026-04-15
We dogfood the opt-in Reliability Loop and use our own eval system to gate tuned-adapter releases. Here is a case where it actually worked.
3 min read
Why we didn't build a separate trace pipeline
2026-03-22
The observability design decision that eliminated the export-label-upload cycle — and why it means corrections feed the optimization loop without a pipeline.
4 min read
How Atlas picks a model
2026-02-01
Atlas is the optimization layer, not a model. Here's what it conditions on per call to serve the cheapest path that holds quality — and how that sharpens as your operation accumulates traffic.
3 min read