How it works / request path

How a request reaches an approved model

Your team sets the allowed models, providers, location, budget and fallback rules. Gyde checks each route against those rules, serves on approved capacity and records the outcome. If no approved path is available, the request can fail.

The managed setup has been tested in one geography. Your workload and environment need their own pilot checks.

Review deployment controls See test limits

01Request path

The request follows an approved route.

In the current prototype, the ARBR gateway (open source and OpenAI-compatible) and the router check the policy before sending a request to a model and host. ARBR applies approved-model routing, budgets, rate limits and kill switches, and meters each request for audit. The response returns with a record of the route and outcome. The router's decision log holds no request content by default; log policy for the rest of the stack is set per deployment.

01Client request
02Gateway and router
03Policy decision
04Approved model and host
05Response and metadata trace

02Before serving

Model qualification still needs a human review.

01

Define the workload

Agree on the required output, test cases, unacceptable errors, response time, budget, data location, permitted providers and fallback.

02

Qualify candidates

Compare a model trained for the task, the current baseline and any permitted fallback on the same cases. People review the results; qualification is not self-service.

See evaluation consulting
03

Select a permitted path

Gyde checks prices on 9 Indian GPU sources every 10 minutes and chooses by total job cost. A cheaper GPU is excluded when it violates location, approval, hardware compatibility or recovery requirements. Some placement controls are deployed and still in live acceptance.

03Operational decisions

A replacement route must still be allowed.

01

Serve privately

The tested prototype uses an edge gateway, router, private network and GPU node agent. Engine ports are not public. Prompts still traverse the approved gateway and provider path.

02

Handle interruption

An unhealthy host can be removed from routing. An approved fallback or replacement is used only when policy permits. If none is permitted, requests may fail.

03

Review the record

See which route was used, whether the request succeeded, what happened to capacity and the available cost evidence. Versions are recorded per test run. A full billing ledger remains planned.

04Policy outcomes

Blocked and unavailable are real outcomes.

01

Allowed

Use an approved model, location, provider and capacity.

02

Blocked

Do not route a request when policy prohibits the available path.

03

Unavailable

Report failure when no approved healthy path exists; do not imply zero downtime.

?Buyer questions

Questions to settle before a pilot.

What happens when a GPU fails?

The tested system detects an unhealthy host, removes it from routing, and looks for an approved standby or replacement. The request can fail during interruption, and recovery time depends on available capacity and policy. The controlled tests are described on the evidence page.

Can a workload prohibit fallback?

Yes. Fallback permissions are part of the workload policy. If the only healthy model or provider is outside the allowed set, the system must block that route even if it appears faster or cheaper.

How is hardware compatibility checked?

Candidate capacity is assessed against the model's hardware and serving requirements before placement. Some placement controls are deployed and still in live acceptance. The customer pilot must verify the actual configuration and behaviour.

Does the control plane sit in the request path?

The gateway and router serve the request path. Control functions manage policy, placement and capacity decisions, while metadata traces record the route and outcome for review. Some placement controls are deployed and still in live acceptance. The exact network boundary is mapped for each deployment.

Gyde Inference pilot

Start with a workload you can measure.

Describe the task, data boundary and decision you need to make. We will scope the test and operating conditions together.

Discuss a workload Do not submit sensitive records through the marketing form.