AI providers move. The product cannot.
Outages are routine, the best model changes every few months, and every provider works differently. None of that should reach the product team.
| What | Before | After |
|---|---|---|
| A provider fails | The product fails | Work moves to the next one |
| Switching AI models | Rebuild the product | Changed behind the scenes |
| Adding a capability | A new project | One new field |
| Provider quirks | Leak into the product | Handled in one place |
| Stuck requests | Hang indefinitely | Time limit, then reroute |
| Tracking a request | Guesswork | Traceable end to end |
How we did it
- 01
One simple agreement
The product sends one message describing what it has and wants, and gets one answer back.
- 02
Every provider behind one door
Each provider’s quirks are handled in its own adapter, never in the product.
- 03
Health decides the work
Providers that fail are sidelined, and win work back when they recover.
- 04
Five generations, no disruption
The video model changed five times in a year. The product never changed.
One moment from the work
Built for bad days“What happens when a provider goes down?”
Nothing the customer sees. A failed attempt costs the time it took, then the work moves to the next healthy provider.
- Delivery routes
- 25
- Model generations
- 5
- Product changes
- 0
What any team can take from this
Put one door in front of vendors
Your product should not know which vendor did the work.
Let health decide
Never switch a vendor off by hand. Let the scores do it.
Describe what you have
The request says what it has; the system decides the job.