Custom models
for your AI agents.
Models trained for your production AI agents on one API, improved by continual learning.

From agent workload
to specialist model.
Scope the agent workload
We review the steps your agents repeat, their inputs and expected outputs, and your current quality, latency and cost.
You receive: A defined agent workload and agreed quality bar.
Build and evaluate
We turn representative agent requests and expected responses into training and test sets, then evaluate the specialist against your current model.
You receive: Your specialist model and workload evaluation results.
Deploy and keep learning
We agree how the specialist connects to your agents, where it runs and who operates it. Once live, it learns continually from your agents’ traffic: each new version trains in the background and is scored against your evaluations.
You receive: A live specialist that improves with every release you approve.
Evidence from
a defined task.
One model, measured on one meeting-summarisation benchmark (QMSum). Your agent workloads get their own evaluation.
99%
lower cost than Claude Opus 5
and 96% lower than GPT-5.6 Luna
+18%
better after training
ROUGE-1, same model before and after
23×
faster than Claude Opus 5
and 6× faster than GPT-5.6 Luna
Method and limitations: Read the paper
Your model,
under your control.
Your specialist learns continually from your agents’ traffic, and you decide what it learns. You approve the training examples and every release. If a new version falls short, the current model stays in service.
- Deployment
- Ertas-managed, in your cloud or on-premise.
- Ownership
- Runs on Ertas by default. Ownership and handover come with a minimum commitment.
Start with one
agent workload.
Bring one agent workflow, a few example requests and your current costs. We’ll show where a specialist model fits.