How to Use Jev AI
Access is waitlisted and the query model is unlike any LLM you have used. Here is how to get in and how to ask your first question.
Five steps to your first Jev AI answer
Join the waitlist
Access is gated. Sign up at typesafe.ai and wait for the invitation email — developers are pulled off the list over time rather than all at once.
Define your schema first
Unlike an LLM, you do not write a prompt and hope. You declare the shape of the answer: the option list for a Choice, the scale for a Score, or the binary for a Noul.
Send the question with context
Provide the input material and the typed question. The model returns a value inside your schema plus a calibrated confidence figure.
Set a confidence threshold
Decide in your own code what confidence justifies acting automatically. A common starting point is a 0.5 floor for genuinely uncertain answers, raised as the stakes rise.
Calibrate on your own data
Run it against a labelled sample, plot confidence against accuracy, and move your thresholds to match what you observe rather than what the docs suggest.
Unlearning the prompt before you use Jev AI
The hardest part of picking this up is not technical. It is that two years of habit tell you to write instructions in English, ask nicely for JSON, and add a few examples for good measure. None of that applies to Jev AI. There is no system prompt to tune, no temperature to lower, no "respond only with valid JSON" incantation to append. The schema is the instruction.
Which means the work moves upstream. Instead of iterating on wording, you iterate on the question itself: are these really the right categories, is this scale the right granularity, should this be one decision or two chained ones. Teams who have used structured output on a language model will recognise the shape of that work — it is the part that was always doing the real lifting, with the prompt engineering stripped away.
Choosing between the three Jev AI shapes
Choice
Routing, classification, triage, picking a category or an action. Example: Which department should handle this support ticket?
Score
Severity ranking, quality grading, priority assignment. Example: How urgent is this incident, on a 1–5 scale?
Noul
Binary gates, eligibility checks, policy and safety decisions. Example: Does this submission violate the content policy?
When a problem seems to need two of them, it usually is two questions. A moderation pipeline might run a Noul gate first — is this in violation at all — and only then a Choice to say which policy applies. Two cheap calls chained is almost always clearer than one overloaded schema, and at this price point the second call costs nothing worth counting.
Calibrating Jev AI before you trust it
Every Jev AI answer arrives with a confidence figure, and the temptation is to take that number at face value on day one. Do not. Pull two or three hundred examples you already have labels for, run them through, and plot confidence against whether the answer was correct. What you want to see is that the cases the model rated 0.9 really are right about nine times in ten.
That curve is what your thresholds should be built on — not a number from a blog post. A reasonable starting posture is to auto-act above the level where observed accuracy clears whatever your process demands, route the band below it to a human, and log everything in between so the boundary can move as you learn more. On low-stakes routing the bar can sit low. On anything that touches money, access, or safety, set it high and keep a person in the loop regardless.
The other thing worth doing early is measuring what Jev AI is replacing. If a rules engine currently handles 60% of your cases correctly and dumps the rest on a queue, that is the number to beat, and it is often a lower bar than people assume. Conversely, if a human team is at 97%, a model at 91% is not a replacement — it is a pre-filter that makes the human faster.
Finally, plan for Jev AI to be wrong occasionally in ways that are valid. A misrouted ticket is a normal operational event. Build the same correction path you would build for a new human hire: an easy way to flag it, a record of what was flagged, and a periodic look at whether the mistakes cluster somewhere your schema could be fixed.