Decisions without words.

Ineffable answers typed questions about a piece of text: which team, how urgent, will they leave. Every answer comes back as calibrated probabilities in milliseconds, with no generated text to parse or doubt. Train it on your own labels in minutes.

Request access Read the API
Real decisions from our models, shown one after another. For example, for the message "billed twice, refund please or we cancel", the model nhsd/tickets@v1 answered three questions in 14 ms: department billing at 100%; urgency critical at 30.4%, below the threshold, so it escalates; churn yes at 74.3%.

Three kinds of question cover most decisions.

You describe the options in plain language. The model reads the text against each option and returns a probability for every one of them, so you always know how sure it is.

Choice

Pick one of your labelled options: a department, an intent, a category.

"dept": {"type": "choice",
  "instructions": "Which team?",
  "criteria": {"billing": "refunds",
               "tech": "bugs and outages"}}

Score

Place the text on a scale you define, level by level: urgency, risk, sentiment.

"urgency": {"type": "score",
  "instructions": "How urgent?",
  "criteria": ["not urgent", "normal",
               "urgent", "critical"]}

Yes or no

Decide whether a statement holds: needs review, threatens to leave, is spam.

"churn": {"type": "noul",
  "instructions":
    "The customer threatens to leave."}

Measured on our own hardware, not promised.

14 ms
to answer three questions about one text
on a single RTX 4070, model already loaded
36 ms
p95 per question with 20 clients at once
196 requests a second; requests share GPU passes
89%
accuracy after a two-minute fine-tune, up from 44%
eight intents in four languages, 215 messages written by hand
95%
right when it is at least 90% sure
and at that threshold it still handles 84% of messages itself

The accuracy figures come from the first production workload: routing messages to a strength coach. Base model 44%; fine-tuned on 3,579 synthetic messages, 89% on messages it never saw. Your numbers depend on your data; every fine-tune reports them before you deploy.

From labelled examples to a live model, in one loop.

  1. Upload examples

    A JSONL file of texts, questions and the right answers. Every record is checked before anything is stored.

  2. Fine-tune

    Minutes on a GPU. The report shows accuracy, calibration and how it compares with the model you run today.

  3. Deploy a version

    One click or one call. Callers keep using the same name; rolling back is just as quick.

  4. Label what it decides

    Every answer is logged with its request. Check the uncertain ones, label them, and they are your next training set.

Pick the confidence threshold from the report: above it the model decides, below it you escalate to a person or a larger model.

Built for the decisions you make thousands of times a day.

Where it fits

  • Routing tickets, emails and chat messages
  • A fast first pass in front of an LLM agent, which escalates only what it is unsure about
  • Guardrails: is this request allowed, does this answer need review
  • Triage scores for invoices, incidents and traces

How you use it

  • POST /v1/systemone, compatible with TypeSafe Jev clients
  • An MCP server, so Claude can train, check and deploy models for you
  • A command line tool and a dashboard for your team
  • Your weights stay yours: download any checkpoint, upload ones you trained elsewhere

Request access.

We are opening Ineffable to a small number of teams. Tell us what you want to decide and we will set you up.

Thanks. We will be in touch.

That did not work; email hello@ineffable.nl instead.