A well-being scoring tool that helps schools notice the students who need support

Client Industry

Education technology · schools & pastoral care

Key Challenge

Spotting struggling students before grades reveal it

Core Tech Stack

Deep learning · Survey analytics · Time series · Python

Deployment Model

Scored history with human-owned escalation

SoftmaxAI developed a student well-being scoring tool which enables pastoral staff to identify a student who is having difficulties before such issues appear in the results. The tool combines structured questionnaires designed to assess a student's mood and level of engagement with signals that the learning platform already has — for example, the material a student selects to study and how frequently they ask for support — and then evaluates the student across physical, mental, and emotional aspects. The scores are kept as a record of trends rather than as a final judgment, so that staff can observe a trend; if this trend goes beyond an agreed threshold, the tool raises it with a member of staff. It suggests a conversation, without making any determination about the child.

• The signal arrives too late. Generally by the time there is a drop in the assessment results, the period when support could have most easily been provided has already gone by.
• Staff attention does not scale. Noticing a change in one student among hundreds depends on the right teacher happening to spot it — reliable for some students, invisible for others.
• A single reading means nothing. It's entirely possible to have a bad week; a tool which responds to a single low response will produce noise — all that's important is whether or not the direction of travel has changed.
• The wrong design becomes surveillance. When measuring well-being it is easy to end up with a system that monitors children. The inputs had to be legitimate and obtained with the consent of the relevant parties, and the output had to be directed at informing a human rather than replacing that person.

The tool operates on a left-to-right pipeline whereby the responses given in the questionnaire when consent is obtained and the signals indicating platform engagement are combined to form features for each student, then scored against various well-being dimensions, tracked as a time series, and only passed on to staff when a sustained trend goes beyond an agreed threshold.

• Boundaries written into the spec. Prior to carrying out the modeling, we decided which inputs were legitimate, what the score did and did not claim, set the escalation threshold, and established the rule that the tool would never produce an outcome for a student — it would only provide a prompt to a person.
• Consented inputs only. Scores are built from questionnaire responses the student gives and engagement signals the platform already holds with their knowledge — not from monitoring behavior they did not agree to share.
• Trends, not snapshots. The deep-learning models evaluate each well-being dimension on a continuous scale and consider the trend over time, meaning that one low response is not enough to trigger an intervention.
• Human-owned escalation. The client’s educators were responsible for setting the thresholds, choosing the alert wording, and handling the pastoral response — verifying them against real-life situations to ensure that an alert always meant "have a conversation".

• Earlier visibility than results allow. A change in engagement and mood can be detected while support is still easy to provide, rather than only after a period of decreasing marks.
• Consistent attention across every student. Every student is scored on the same basis, so being noticed no longer depends on which teacher happens to be paying attention.
• An evidence base for difficult conversations. Pastoral reviews and parent meetings start from a documented history across well-being dimensions instead of impressions recalled after the fact.
• Alerts that mean something. Threshold-crossing on a sustained trend — not a single response — is what reaches staff, which is what keeps the alerts worth reading.
• Decisions stay with people. The tool surfaces and evidences; educators interpret and act. No score produces an automatic consequence for a student.