Skip to content

Artificial Intuition

Decision models that keep getting better at your work.

Open the consoleSelf-host with open-server (opens in a new tab)

Hosted: VAKYN MAX, with an API key. Your servers: open-server, free and MIT.

Who it's for

For teams that make the same decision thousands of times a day.

  • Support inbox

    Every ticket routed as it comes in. The unsure ones go to a person.

    choiceWhich team should get this ticket?

  • Security alerts

    Every alert triaged the moment it arrives. The unsure ones go to an analyst.

    scoreHow urgent is this alert?

  • New certificates

    About 6 million new TLS certificates a day. VAKYN checks every one as it appears.

    noulIs this domain pretending to be a bank?

    Volume: public Certificate Transparency statistics.

  • Expense reports

    Every report checked against the policy before finance pays it.

    choiceApprove, flag or reject this expense report?

  • Hiring

    Every CV checked against every skill the role asks for.

    noulDoes this CV show Kubernetes experience?

  • Reviews

    Every review checked before it goes live on the product page.

    noulIs this product review fake?

We run it

VAKYN MAX

Our hosted version, with our biggest model. Send a document and your questions with an API key; nothing to install.

Open the console
You run it

open-server

The open models and a free server on your own machines. Already on a Jev-compatible API? Change the base URL.

open-server on GitHub (opens in a new tab)
Why VAKYN

AI that answers instead of talking.

You ask a typed question about a document: yes or no, pick one, or a score. VAKYN returns the answer and how sure it is.

  1. 01

    Gets better at your work.

    The VAKYN loop finds where it fails on your cases, trains exactly there, and proves the fix. Again whenever you need it.

    997occupations
    in 22 domains, covered by the training
  2. 02

    Sovereign.

    It runs on your servers, in your country, under your rules, and your data never leaves. Or we run it for you on VAKYN MAX.

    Apache-2.0
    open model weights; the server is MIT
  3. 03

    Fast and cheap at any scale.

    Milliseconds per decision on one graphics card. On your own servers there is no request limit.

    16.6ms
    median, one question on a short ticket (179 tokens); open-server, VAKYN-4B Q8_0, one RTX 5090
One question, two answers

An LLM writes a paragraph. VAKYN picks an option.

The same expense report and the same question. VAKYN returns one option, a probability for every option, and how sure it is.

expense report, as JSON

Ines Varga, field engineer

Customer site visit in Lyon, 6 October 2026

  • Flight, Brussels to Lyon and back, business class, booked privately1240no receipt
  • Hotel, three nights for a one-day visit1260no receipt
  • Dinner, alone, two bottles of wine310no receipt
  • TotalEUR 2,810
The policy sent with it
Rail
Second class, booked through the travel desk
Hotel
Up to EUR 180 a night
Meals
Up to EUR 60 a day while travelling, no alcohol
Receipts
A receipt for every line over EUR 25

choice

What should finance do with this expense report?

approve · flag · reject

An LLMillustrative

This report breaks the policy in several places. The flight is business class and was booked privately, the hotel covers three nights for a one-day visit, the dinner includes two bottles of wine, and no line has a receipt. I would reject it, or at least flag it for a closer look.

Options a parser finds
flag, reject
How sure, as a number
none

Written by us to show the shape of a chat reply. Not a measured output.

VAKYNreal answer
reject82% sure
  • approve0%
  • flag12%
  • reject88%
At 70% sure or more, your code acts: send it back to the employee to fix.

Captured 2026-10-11 from the open model. Nothing edited.

No streaming

Nothing to wait for. Nothing to parse.

An LLM sends its reply a few words at a time, and your code can read the decision only after the last word, if it can find it. VAKYN sends one typed object, complete.

An LLM, streamingwaiting
This report breaks the policy in several places. The flight is business class and was booked privately, the hotel covers three nights for a one-day visit, the dinner includes two bottles of wine, and no line has a receipt. I would reject it, or at least flag it for a closer look.
  1. Wait for the last word
  2. Search the text for an option
  3. Found flag and reject. Which one did it mean? No number says how sure.

Illustrative reply, written by us.

VAKYN, one responsewaiting
{
  "answers": {
    "decision": {
      "choice": "reject",
      "confidence": 0.82,
      "probabilities": {
        "approve": 0,
        "flag": 0.12,
        "reject": 0.88
      }
    }
  }
}
Read choice and confidence straight from the object. The same fields every time.

Real answer to the clear expense report above.

Time per decision
16.6ms
median for one question on a short ticket (179 tokens); open-server, VAKYN-4B Q8_0, one RTX 5090
Animation not to scale.
In a workflow

Act on the sure cases. Send the rest to a person.

In an automated workflow the number decides where each case goes. Pick the threshold on your own cases; everything under it goes to a person.

decide.py
answer = reply["answers"]["decision"]

if answer["confidence"] >= 0.70:
    act(answer["choice"])  # sure
else:
    send_to_person(case)   # unsure
An LLM in the same place

“I'm confident it can be approved.”

A chat model sounds just as sure when it is wrong. “I'm confident” is part of the text it writes, not a measurement, so there is no number to set a threshold on. You either trust every case or check every one.

Illustrative reply to the borderline report above, written by us.

Six cases, routed at 70%3 acted on3 to a person

  • Agent refund, EUR 900

    Battery returned sealed and checked in by the warehouse

    allow

    100% sure

    Code acts

  • Review of an espresso machine

    One star: stopped heating in week five

    quality

    97% sure

    Code acts

  • Expense report, EUR 2,810

    Private business-class flight, no receipts

    reject

    82% sure

    Code acts

  • Expense report, EUR 498

    EUR 27 taxi without a receipt, card reader broken

    reject

    65% sure

    Person checks

  • Eight job applications

    Night shift lead, several close candidates

    application_3

    49% sure

    Person checks

  • Review of a kettle

    Boils fine, arrived late in a crushed box

    quality

    23% sure

    Person checks

Real VAKYN answers, captured 2026-10-11 from the open model: one choice question per case. The line on each bar is the 70% threshold.

From the blog

Notes on deciding at volume.

All posts

Your work is not a benchmark

Models are tuned to raise general scores. The only score that predicts how a model will do on your work is a score on your work.

3 min read

Read the post

Show us your hardest decision.

We'll show you where VAKYN gets it right on your cases, where it gets it wrong, and how the VAKYN loop fixes that.

Open the consoleSelf-host with open-server (opens in a new tab)

vali@vakyn.com