SovrGPT Decision Engine
Decisions via API, in milliseconds.
The Decision Engine reads a text or system state and rates it against your rules: which team, how urgent, whether a human should look. Your software gets probabilities and a recommendation back and decides for itself what runs automatically.
Incoming message
“Since this morning I cannot log in. My September invoice was also charged twice – please check this urgently.”
Evaluation
Recommendation: review – no automatic reply
Sample measurement without double check · 397 ms · no generated text
At a glance
Three formats
Choice among up to 26 options, yes/no, or a level on a scale.
In milliseconds
The model writes no text, so a question takes milliseconds. Several questions about the same text run in parallel.
A human when unsure
Uncertain cases are flagged for review. Every answer states the model version and where it was processed; nothing is stored.
Germany or EU
v1 runs in Germany, v2 in the EU. Via the API, in the chat as the "Decide" tool, and over MCP.
Ready-made templates
- Ticket routing
- Inbox classification
- Customer feedback analysis
- Lead qualification
- Public-sector inbox
- Application complete?
- Pre-publication check
- Agent routing
- Grounding check
- Agent step gate
Three formats
Choice
One of up to 26 options – team, category, next action. With a probability for every option.
Yes/no
Does a statement hold? Review needed, complaint, escalation – as a probability, not a gut feeling.
Ordered score
One level of a scale: urgency, relevance, quality. With an expected value between the levels.
Why it is fast
A chat answer is written word by word. For a decision the model reads the text and options once, and we read off how likely each option is. Several questions about the same text run in parallel. Choice questions are also checked with reordered options, in parallel as well.
Our measurement (warm, ticket text, four options, without double check):
- v1 (Germany): 65 ms per question · four questions at once 123 ms
- v2 (EU): 198 ms per question · four questions at once 397 ms
An image can be the state too. So far we have only tested this with a small test image; we have not yet measured how long photos or scans take.
A chat answer to the same text takes 5 to 30 seconds. These are measurements, not a guarantee.
One API access, clear rules
- Your application
- SovrGPT API: identity, tenant, scopes, limits
- Data centres in Germany or the EU
- Rating back, your application decides
Every answer states which region and model version did the computing. The core path calls no model API of a US provider.
Next, we are training and calibrating specialised decision models based on open base models, with documented tests for each task.
Transparency
Even a cleanly structured answer can be wrong. That is why every answer states the model version and calibration status, and uncertain cases go to a human for review. Binding approvals stay in your application.
We do not store the text, questions or answers. Billing is by tokens. The processing location is in every answer.
Operating models
Decision API
Shared inference pool with API access, quotas, docs and playground. Ready to use.
Dedicated Decision
Dedicated operation with contractually defined region and performance. On request.
Custom Decision
Data preparation, training, evaluation and maintenance of a model for your task. On request.
For developers
API-key (Bearer) authentication, one endpoint, JSON in, JSON out.
curl https://sovrgpt.com/api/v1/decisions \
-H "Authorization: Bearer $SOVR_KEY" -H "content-type: application/json" \
-d '{ "preset": "ticket-routing-v1", "state": "…" }'Go to the API documentation Prices on request. The probabilities are not calibrated yet. We keep measuring and will let you know when limits or model versions change.