JITM Evergreen is live: integrate once, stay right

Every model starts going stale the day it ships. JITM Evergreen keeps your prediction endpoint in step with the real world: same URL, no redeploys, and nothing changes without your approval.

jitm evergreenOne endpoint, many generations

Your app

or your agent, spreadsheet, CRM

Integrated once

Same URL, forever

/api/predict/model

No redeploy. No new keys. No code change.

Generation 1Retired
Generation 2Retired
Generation 3 Serving
ChallengerBeing tested

Your app talks to one endpoint. Behind it, a better model can take over without you touching a line of code.

Every model is a snapshot of the past. The day it goes live, the world starts moving away from it: prices shift, customers change habits, a competitor launches. Usually nobody notices until the numbers look off, and the fix is a whole new project. JITM Evergreen is our answer, and it is now live in production.

Integrate once

Wiring a model into your app, CRM or AI agent is the expensive part, so you only do it once. Your endpoint URL never changes. When a better model earns its place, it simply starts answering behind the same address. No redeploy, no new keys, no code change.

Drift is expected. We deal with it.

Models going out of date has a name: drift. Most tools treat it as an emergency. We treat it as routine. Send back real outcomes as you learn them (the actual sales, the settled price, the resolved ticket) and Evergreen tracks how the live model is really doing. It knows a bad day from a real shift, and when the shift is real, it trains a challenger on the new reality.

jitm evergreenLive error, checked against realityIllustrative

Live error (lower is better) Training baseline Drifting

A one-off spike is ignored. A sustained climb is flagged, a challenger is promoted, and error drops back to normal.

Better models, never behind your back

A challenger has to win a fair fight: both models are scored on the same fresh outcomes that neither trained on. If it wins, it waits for you to press Promote. Every generation is logged, so you can always see why the model changed.

jitm evergreenChampion vs challengerIllustrative

Scored on the same 120 fresh outcomes. Neither model trained on them.

Champion

Generation 2, serving today

14.2avg error

Trained before the market moved

Challenger

Trained on what happened since

10.6avg error

25% lower error on the same rows

Awaiting your approval

Already running on real data

The first model on Evergreen is Morrowatt, our day-ahead power price forecaster for Germany-Luxembourg and Spain. The auction settles every day, so the truth arrives within 24 hours and the loop turns in days, not months.

For the geeks

Every prediction now returns a prediction_id; post the outcome back when you know it. That is the whole integration. Drift is measured on real accuracy against real outcomes, not guessed from the inputs, and a challenger only takes over if it beats the champion on fresh outcomes neither model has seen.

POST /api/outcomes/{prediction_id}
{ "outcome": 495 }

Get started

JITM Evergreen is available on the Scale and Enterprise plans. If you have a prediction problem where the world keeps moving, we would love to hear about it.

Build a model