A decision model, not a chat model

Uprelic Strom: Every decision in your business. Under 200 ms each.

Ask Strom in any modality. In plain language, about text, images or video. No training data or fine-tuning needed.

Get answers your code can act on. A clear yes or no, a choice or a score, ready to use in your code.

Trained on your industry’s decisions. Fine-tuned on real B2B work in 14 business domains and many more, from invoices and contracts to machine logs and web agents. Processed in the EU.

Read the guide

Workflows: Chain decisions.

Real processes are a chain of small decisions. Strom takes about 90 to 430 ms per request in the model, depending on how much it reads. With the network round trip, the workflows below finish in about a second, several times faster than a chat model and at a fraction of its cost.

Many questions, one request

Ask everything you need to know about the same item at once: language, team and urgency come back together.

Branch on every answer

A yes, a choice or a score sends the item to its next request, with more to read if it needs it, in your own code, where you can test and change it.

Run it on everything

Fast and cheap enough for every ticket, invoice and log line, not just a sample.

Decision layer: Let your LLM write. Let Strom decide.

Strom doesn’t replace your chat model. It takes the small decisions around it, so the LLM only runs where you need words.

Smart if-statements
Replace brittle rules and regexes with a question in plain language.
Guardrails for your LLM
Check what goes into an LLM and what comes out, before anything is sent or run.
Every row, not a sample
Score every ticket, log line, review or frame in a dataset for cents.
Real time
Fast enough for user-facing features, live video and agents that act step by step.

No hallucination: It can’t make things up.

Strom never writes free text. Every answer is one of the options you defined, so an invented fact, category or promise can’t appear in its output.

Only your options
A yes or no, one of your choices, or a point on your scale. Strom scores exactly those options and nothing else.
Says when it’s unsure
Every answer carries a probability. A low one is your cue to send the case to a person.
Wrong, never invented
Strom can pick the wrong option, like any model. It can’t make up a new one, and your threshold catches the doubtful cases.

Speed and price: Faster than a blink. Priced like a rounding error.

A decision model answers in one forward pass, not a paragraph of reasoning. That’s why most requests take under 200 ms in the model and cost cents per thousand answers. End to end, that is 5 to 10 times faster than a chat model.

$0.042 per million input tokens, output free

Credit where it’s due: The idea came from Jev.

TypeSafe’s Jev showed that everyday decisions deserve a model of their own: typed questions in, calibrated probabilities out, in milliseconds instead of seconds. It’s the System One of software, the fast, intuitive half of thinking. We loved it the first time we used it.

So we kept the shape: Strom speaks the same request format. Then we built our own model, for what European businesses asked us for.

Why we built our own: Same idea. Different priorities.

We wanted a decision model we could run in Europe, trust with high-stakes calls, point at images and video, and teach new languages and domains. So we trained one.

Text, images and video. Natively.

One model reads all three, with no OCR, captioning or vision pipeline in front of it. Put text and several images in one request, or send a video frame by frame and watch the answer change as the scene does.

Trained on the decisions businesses actually make.

Not trivia, not chat. Every domain we train Strom on comes from what businesses need every day: routing tickets, checking invoices, reading contracts, triaging machine logs, steering web agents. It works on your processes from day one, with no training data of your own.

Speaks Europe. Learns your domain.

Every training cycle makes it better in German, French, Italian, Spanish, Polish and more. Next up are specialist versions for web agents, product catalogs and documents.

Processed on GPUs in Europe.

Support emails, product photos and camera feeds are processed on GPUs in the EU, run for us by European cloud providers. No request is forwarded to a third-party AI provider.

Video, natively: Ask a video a question. Every frame answers.

Send frames as they arrive and the same question gets a fresh probability each time. There’s no vision pipeline to glue together, just a question in plain language.

See it decide: One request. As many questions as you need.

Ask up to 256 questions about the same input. Each answer comes back with a known shape, ready to route, score or review in code.

No fine-tuning: Just say what you want.

A classifier used to be a project: collect labels, train, evaluate, deploy, and start over when the categories change. Now it’s a sentence.

Add a category by adding a line. Change a policy by changing a sentence.

Questions: What people ask us.

How is this different from JSON mode or structured outputs?

JSON mode constrains the text a chat model writes. It still writes token by token, can still break the schema, and gives no calibrated probability. Strom never writes: it scores your options, and our API builds the JSON from the scores.

Can Strom be wrong?

Yes. It can pick the wrong option, like any model. It can’t invent a new one, and every answer carries a probability, so you can act automatically on confident answers and send uncertain ones to a person.

When should I still use an LLM?

For open-ended work: writing replies, summarising, planning, code. Use Strom for the decisions around that work, and your LLM runs far less often.

Does it work with images and video?

Yes. Send up to eight images per request and ask about them, or send a video frame by frame and get a fresh answer for every frame.

Where is my data processed?

On GPU servers in the EU, operated for us by a European cloud provider. No request goes to a third-party AI provider. Request contents are deleted after 90 days and never used to train any model.

I already use Jev. Can I switch?

Strom accepts the same request format: a state and typed questions. Point your client at our API and use your Uprelic key.

Who is behind Strom?

Uprelic GmbH, a team in Berlin that builds AI for everyday business work. Strom grew out of the decision steps of our own AI agents.

Ask your first question.

$0.042 per million input tokens. Output tokens are free. Pay as you go, no subscription.

Coding agents: everything about Strom (what it is, when to use it, the API, the Python SDK, pricing and limits) is in one Markdown file at https://platform.uprelic.com/skill.md. It doubles as an agent skill. For a hands-on tour with live calls, see getting-started-with-strom.md.