Quality floor, latency and energy ceilings
Joulis routes each AI request to the model that fits your quality floor, latency ceiling and energy ceiling.
Policysupport-summaries
Routing within budget
Request 118 of 200 routed to Gannet-S
Gannet-L held for 50 of 200
90/100 batch
Floor at least 89 out of 100.
1.4s p95
Ceiling at most 1.8 seconds p95.
50Wh / 200
Ceiling at most 60 watt hours per 200 requests, estimated.
Quality floor
The lowest score a route may return.
Latency ceiling
The slowest p95 the workload accepts.
Energy ceiling
Estimated accelerator energy per workload.
Set quality alone and every request goes to the largest model. Set energy alone and quality drifts. Together, a route has to earn its place.
or Joulis reports the exception before it runs.
89/100
Without it, the cheapest route wins and summaries slip. Scored on a fixed rubric.
1.8s p95
Without it, a quality-only policy waits on the largest model. Measured at the gateway.
60Wh / 200
Without it, the baseline sends all 200 to Gannet-L: 80 Wh est., over the ceiling. Facility overhead not included.
One route the team chose in advance, used only when nothing else fits.
Dispatch-200 benchmark
200 support ticket summaries, the same limits for both. Latency is measured, scores use a fixed rubric, energy is estimated.
Policy support-summaries
Same workload, same limits, two ways to route it.
Over the energy ceiling
All 200 requests to Gannet-L
Meets the policy
150 to Gannet-S, 50 to Gannet-L
37.5%
Lower estimated batch energy for this test, with both scores above the floor.
Estimates cover accelerator energy only and leave out facility overhead.
No route meets this policy
Reported before execution. Nothing ran.
support-summaries-fastExample policy≥ 89/100 · ≤ 1.2 s p95 · ≤ 60 Wh
| Candidate | Score | p95 | Energy est. |
|---|---|---|---|
| Gannet-L | 92 | 1.6 s | 80 Wh |
| Budgeted mix | 90 | 1.4 s | 50 Wh |
A policy never quietly breaks its own budget. If no available route clears all three, Joulis reports the exception before the request runs.
Every candidate route, the budget it fails and by how much, in the response your service already reads.
A policy can point to one fallback route. It runs only when the team chose it in advance.
The request returns unexecuted, so the team can relax a budget or hold the rollout.
Send the policy id with the request. Joulis checks every eligible route against it, runs the one that fits and tells you what it cost.
Add a policy id to the request you already make. Nothing else in your client changes.
Each eligible model is scored against the quality floor, latency ceiling and energy ceiling before anything runs.
The route that clears all three budgets runs. When none does, the exception returns first.
Every response carries the route, measured latency and an energy estimate, ready for your own reporting.
POST /v1/route
policy: support-summaries
route: gannet-s
Routing, policies and reporting in one monthly fee. Your model providers bill inference directly.
Included
Billed by Joulis€99/month
The gateway fee, €99 a month for 100,000 requests.
Billed by your providersper call
Inference charges for every model call, at their rates.
Energy figures in reports are estimates of accelerator energy and leave out facility overhead.
See how we estimateWhere each number comes from, what it leaves out, and what you pay for.
Read the reporting methodologyLatency is timed at the gateway for every request. Accelerator energy is not metered per request, so Joulis estimates it from the hardware, its utilization and the run time, and states each assumption in the methodology.
It covers estimated accelerator energy only. Facility overhead, such as cooling and power conversion, is not included, and every figure says so next to the number.
Each output is scored against a fixed rubric that stays the same for every policy in a comparison, so a 90 and a 92 mean the same thing on both sides.
Joulis reports every request that exceeds a budget. When no route can meet a policy at all, the exception comes back before the request runs.
No. The gateway fee covers routing, policies and reporting. Inference charges stay with your model providers and are billed separately.
Bring one workload. We will run it under your policy and show you what each route delivers and consumes.
