Use Cases

Turning “off sick tonight” into an updated nurse roster

Absence messages read, checked against the ward's staff and skills, and turned into roster changes — with a manager pulled in only when the message is unclear

Illustration of a hospital ward with beds, care staff and digital monitoring panels

USE CASE / HEALTHCARE ROSTERING

Published product metrics

Semantic alignment accuracy

99.7%

Correct response chosen, planner + GenOR-Twin vs alone

98.0% vs 79.5%

Hallucination catch rate

88.2%

These figures match the published metrics on the Predictim GenOR-Twin product page — this walkthrough is a composite scenario, not a specific engagement.

The challenge

Absence messages read, checked against the ward's staff and skills, and turned into roster changes — with a manager pulled in only when the message is unclear

Segment: Hospital nurse rostering

The roster lives in a scheduling model; the disruptions live in messages, calls and handover notes. A ward manager has to translate each one by hand, often at night, and either patch the roster locally or rebuild it — without much time to check whether the patch quietly breaks a skill-mix or rest-time rule somewhere else.

The approach

How the deployment was structured

Read the message, propose the constraint

A message like “night nurse on ward 4 off sick tonight” is translated into a constraint the rostering model understands — this person, this shift, unavailable — with a confidence score attached.

Checked before it touches the roster

The proposed constraint is validated against the ward's knowledge graph: the staff member exists on that ward, the shift is in the future, and the change is physically possible. In offline validation, this layer caught 88.2% of the incorrect extractions a language model produced on highly ambiguous messages.

A response sized to the problem

If cover exists within the roster's slack, it's a local repair. If not, or if a critical skill would be missing, the roster is fully re-optimized. If the system isn't confident it understood the message, a manager checks that one interpretation instead of rebuilding the roster.

What changed

Nurse scheduling is one of the six problem types GenOR-Twin was validated on. The figures above come from that published evaluation: 299 of 300 annotated logs translated exactly as experts read them, and planners working with the system picked the right type of response 98.0% of the time, against 79.5% working alone.

Product used