← All examples
Reliability

A shared early-warning system for AI providers, live in a day

Each product had its own private opinion about which AI providers were healthy, and none could share it. A small shared service now keeps one view for all of them.

Client
Broadcast media company
Timeline
One day to live
Role
Lead engineer
New servers
None
01

Good judgment, stuck in one place

The logic that scored providers worked well, but it lived inside one product and forgot everything on restart.

WhatBeforeAfter
Provider healthKnown in one productShared by every product
After a restartForgottenRemembered
Two copies runningDisagreeAgree
Seeing what happensDigging through logsLive dashboards
SecurityNot needed yetLocked down from day one
Before and after
02

How we did it

  1. 01

    Move it, do not change it

    The existing logic was lifted out as it was, so nothing behaved differently.

  2. 02

    Keep decisions local

    Each product still decides instantly and shares what it learns in the background.

  3. 03

    Lock it down

    Sign-in, a closed network, and restricted permissions from the first release.

  4. 04

    Live the same day

    Test environment to production in one day, with no new servers.

03

One moment from the work

Designed for failure

“What if the warning system goes down?”

Every product keeps deciding with what it already knows and catches up when the service returns. The shared view is a bonus, never a dependency.

Idea to live
1 day
New servers
0
Waits per decision
0
04

What any team can take from this

  • Separate the move from the fix

    Move things as they are first. Improve them one decision at a time.

  • Keep the fast path local

    Coordination belongs in the background, not between a customer and an answer.

  • Put a number on every limit

    Every buffer and timeout has a written size.