Synthetische E-Mails und Schreiben einer fiktiven Verwaltung werden klassifiziert, priorisiert und einer Zuständigkeit zugeordnet, jede Entscheidung mit Begründung und Konfidenz. Setzen Sie die Schwelle für automatische Freigabe und korrigieren Sie Fälle, das System lernt mit. Achten Sie darauf, wo der Mensch die Schleife schließen muss.
Synthetic emails and letters of a fictional administration are classified, prioritised and routed, every decision with a justification and confidence. Set the auto-approve threshold and correct cases, the system learns along. Note where a human has to close the loop.
Demo Municipality of Linden Vale · back-office intake · classification, priority and routing with a human in the loop
A synthetic intake of 64 emails and scanned letters is generated from templates with a seeded random generator (same corpus on every reload).
The classifier is deliberately simple and fully inspectable: a multinomial naive Bayes model over word tokens
(P(category | text) ∝ P(category) · Π P(token | category), Laplace-smoothed) decides the category, and a short list of
weighted cue rules (“second reminder”, “statutory deadline”, “no rush” …) decides the priority. Person and street names are masked before tokenisation (the job of a named-entity step in production). Category routing follows a fixed
department table. Confidence is a temperature-scaled softmax of the class scores — a margin, not a guarantee.
The model starts from only 20 labelled seed items. Every approval or correction updates the token counts and cue weights immediately
(online learning); when a priority correction cannot be explained by any known cue, the two most distinctive words of the document are added as
new, weakly weighted cues. The reasoning panel shows exactly which tokens pushed the decision — the same numbers that are highlighted in the document.
The auto-approve threshold models the maturity step from “AI suggests, human decides” (phase 1) to “AI decides routine cases, human handles the rest” (phase 2). Lowering the threshold frees up clerk time but silently accepts the model's mistakes; the “reveal ground truth” switch shows how many auto-approved items were actually wrong — a number you only know here because the corpus is synthetic. In production this is exactly the figure you would have to estimate by sampling.
All documents, persons, companies and addresses are fictional and generated in your browser. Nothing is sent anywhere; the page stores no data.
The point of the demo is traceability: every decision — automatic or human — can be explained by named evidence and reconstructed later.
Keyboard: A approve · C correct · E escalate · J/K next/previous item.