Skip to content
JEV

JEV · intelligent decision & early-warning engine

Changing data in. Decisions and early warnings out.

JEV reads a domain's time-stamped data and finds what is moving. It forecasts where that may lead and scores the risk, then asks fixed questions answered by versioned policies. Warnings, recommendations and actions all come downstream of those decisions, and each one says how confident it is, what kind of confidence that is, and which evidence it rests on.

Create an account Browse the films

The running order

From raw rows to a decision someone can act on.

  1. Observe

    1. Data

      A domain adapter loads time-stamped observations; freshness and quality are checked first.

    2. Signals

      Each change worth a look, with its strength, source and age.

    3. Trends & anomalies

      Theil–Sen slopes with q-values, change points and robust-z breaks.

  2. Anticipate

    1. Forecasts

      Damped-trend models with 80 % intervals, backtested on rolling origins.

    2. Risk

      Likelihood × impact, shrunk by confidence and data quality.

  3. Decide

    1. JEV decisions

      Fixed questions, versioned policies and a declared kind of confidence. It abstains when the evidence is thin.

    2. Warnings & recommendations

      Raised only downstream of a decision, and each carries the decision id.

  4. Act & learn

    1. Actions

      Ranked next steps, each with its expected impact, effort and first move.

    2. Feedback

      Verdicts and outcomes are fed into evaluation, and evaluation checks every stage.

09 feeds back into 01: feedback → evaluation → the next run

Every stage writes its evidence, so a warning can be traced back through the decision, the risk and the forecast to the rows it came from. Nothing past the as-of date is visible to a run, so any month can be replayed without leakage.

Domains

One engine, one adapter per domain.

Everything domain-specific sits behind an adapter. It loads the data, declares which series to build and can add stages of its own. The core engine never imports movie code.

first domain adapter

Movies

MovieLens ratings plus the app's own activity. It adds genre demand, audience lapse, rater behaviour and model governance, and gives each member the drift test and strategy decision behind their recommendations.

Open the movie app

generic dataset adapter

US unemployment

Monthly unemployment rates for the US and a set of states (BLS via FRED, public domain), loaded from a CSV file and a config file. The engine is the same and none of the movie code runs.

Visible in the Intelligence console to administrators.

Movies domain · measured, not claimed

The recommender behind the movie app.

The movie adapter serves a hybrid recommender. Every model in the serving version was evaluated the same way: for each user the latest ratings were held out, the whole catalogue was ranked, and we counted how many of the films they went on to rate 4★ or higher landed in the top 10.

Trending now · covers are typeset from each film's metadata