Many calls to a big LLM are small decisions: which team owns this ticket, which action handles this request. Paste up to 50 of yours and see how many a small model you run yourself answers confidently, on a plain CPU, with your data staying yours.
Privacy: on our server, what you paste answers this request only; it is not logged or stored, and we keep only anonymous counts (visits, runs, clicks). Please don't paste secrets or personal data: it is a public demo. In-browser mode sends nothing anywhere. Limits: 50 lines, 3 checks per 10 minutes.
We have no customer results yet and will publish only ones we can prove. See the full benchmark, including where we lose or 18 recipes with real outputs.
Policies, runbooks, tickets with the team that took them, alerts with whether anyone acted.
Labels come from rules we evaluate or outcomes that really happened, never from another model. You confirm every rule.
An 8-bit model on a plain CPU next to your production. You get a report on your own held-out cases: right and confident, confidently wrong, abstained.
Ten rounds against the clock, then compare with the model on the same ten. Shareable score card. Nothing you do here is stored.
Against a server you run yourself (default port 21118). Real response, trimmed:
curl -s http://127.0.0.1:21118/v1/systemone -d '{
"state": "I was charged twice for invoice 1042",
"questions": {"q": {"type": "choice",
"instructions": "Which team should handle this ticket?",
"criteria": {"web": "frontend or UI", "api": "backend or API",
"billing": "payments and invoices", "docs": "how-to question"}}}}'
{"answers":{"q":{"choice":"billing","confidence":0.945,
"probabilities":{"billing":0.945,"web":0.055}, ...}},"model":"openjevx"}
Question types: noul (yes/no), choice (pick one), score (rating). More in the recipes.
Can your data leave your account?
OpenJevX is open source (Apache-2.0) and Deemwar is its open-source partner. The training book walks from your decisions in a spreadsheet to an 8-bit model you run yourself; The public kit has the scripts and base model; a few pipeline files come from us on request ([email protected]).
We train on our GPUs from your documents and records, and hand you the model, a way to run it, and the report on your own cases. Retrain monthly. Status: pilot, run by hand for the first few teams.
A packaged appliance that trains in your own cloud account so nothing leaves it. Status: early access on AWS (tested in our own AWS account). Not in the Marketplace listing yet. Other clouds later. See the early-access page. Docs
We charge for the hours we save you (data preparation, GPU work, checking), never for the code, and we never host your inference. This demo is marketing; it runs on our machine, your model runs on yours.
Tell us the one decision you want automated. We will tell you honestly whether a small model you run yourself fits, before anything is trained. You keep the model and run it in your own account.