Story
Northwind
▲ Northwind
— Four models, one week, 61% less.
A logistics platform for 9,000 shippers moved four models to Arclight in a week, cut its inference bill by 61% and finally understood its agents.
61%
Lower inference bill
month over month, same traffic
7days
To move four models
with one engineer
2.3×
Faster answers
to the first token, at the 95th
0
Pages at night
since the move, in seven months
Industry
Logistics software
Team
14 engineers, 3 in ML
Region
us-east, eu-west
On Arclight since
March 2026
Plan
Scale
01 The challenge
Northwind’s agents read shipping documents, answer customs questions and reroute parcels when a port closes. In 2025 they ran on three providers, each with its own SDK, its own invoice and its own idea of an outage.
The bill grew faster than the traffic, and when an agent made a wrong call there was no way to see why: the logs stopped at the provider’s door.
02 The move
One engineer moved all four models in a week. Because Arclight speaks the OpenAI format, the change in the code was the base address; the rest of the week went on a fine-tune of Atlas 70B on two years of customs answers.
The document reader moved to Prism Docs on reserved GPUs, the router to Arc-7 Mini on serverless, and every agent run became a trace the team could replay against another model.
“We moved four models to Arclight in a week and cut our inference bill by 61%. The traces alone paid for it: our agents finally make sense.”
Maya Lindqvist
Head of AI, Northwind
03 What runs today
Atlas 70B, fine-tuned on 2 years of customs answers
Prism Docs for bills of lading and invoices
Arc-7 Mini for routing and triage
Guardrails on every answer sent to a shipper
Traces replayed in CI before every release