01
One intake
Applications post to /v1/route. They stop keeping a private map of which model handles which feature.
Loading
Separator
Decision
explanation
reason-1
4.2 ms
Solutions
Put Separator in front of the models your teams already call, and make every handoff typed.
Decision plane
4.2 ms
explanationreason-1
01
Applications post to /v1/route. They stop keeping a private map of which model handles which feature.
02
A team can permit ops-1 and write-1 and refuse vision-1. The refusal is part of the decision, not an error buried downstream.
03
Each decision records the type, the model, the latency, and the parts. Review reads the log, not a blended transcript.
Teams
Business traffic is a pile of applications that each used to keep a private map of models. Separator replaces the map. The application posts to /v1/route and reads the contract.
speech-1, write-1, ops-1
A customer message is split before anyone drafts the reply. The refund, the transcript, and the note are different hops.
ops-1, write-1, reason-1
Closing the books is a workflow. The month note is prose. A question about a missing receipt is a judgment. They do not share a lane.
code-1, reason-1, search-1
A change request stays a diff. The explanation of that diff, and the lookup behind it, leave on their own parts.
search-1, reason-1, code-1, write-1
Incident text fans out. vision-1 stays off the list until someone adds it. A new lane does not become reachable by itself.
Requests
01
Close the month, write the summary, and flag invoices with no receipt.
Fanout. ops-1 runs the close, write-1 drafts the summary, reason-1 names the gaps.
02
Add rate limiting to the public routes and explain the limit to the team.
Fanout. code-1 returns the diff. write-1 writes the note. The explanation does not sit inside the patch.
03
Why did the booking duplicate?
Route. One type, explanation, reason-1. No second model is called to pad the answer.
The application
A feature that used to call one model for every sentence now calls Separator. If the response is a Route, there is one part. If it is a Fanout, there is a list. If it is pending, nothing ran, and the allow-list is the place to look.
Before
Each product kept its own table: this screen calls one model, that job calls another, and a mixed request hoped the model would notice. Review had a transcript, not a hop.
With Separator
One endpoint. The type is assigned first. The team allow-list is applied next. The decision record is what review reads.
Decision record
Prompt text stays out of the record unless the customer enables that logging. Storage region is agreed in writing on an enterprise plan. Downstream model usage is metered apart from the decision.
Engagement
Veloxity is an Xythera company. The work is the integration: which models you already call, which types show up, and which lanes each team is willing to allow.
List the applications, the models they call today, and the requests that currently come back blended.
Write an allow-list per team. A lane that is absent cannot be selected, including a lane we add later.
Point the client at api.veloxity.si/v1/route, or set the model to separator on an OpenAI-compatible base.
Look at types you did not expect. Tighten the list. The log is the review, not a blended transcript.
No. Separator chooses among the lanes you allow. The models stay the ones your teams already use. The new piece is the typed handoff in front of them.
Yes. Support can permit speech-1 and refuse vision-1. Engineering can permit code-1 and refuse speech-1. The allow-list is part of the decision for that call.
In the decision record: type, contract, lane, latency, and parts. The prompt text is omitted unless you enable that logging. Region for the log is an enterprise term, agreed in writing.
No. It is the design target for the routing decision, not for the model’s answer, and not for the consoles on this site. Those consoles run in the browser.
Related: customer support, security, the fabric.