Who
A calibrated dropout probability for every active student, ranked by risk. Not a binary “at risk / not at risk” label.
EO-Churn · Student dropout early warning
Catch the dropout
before it happens.
Which student is about to leave, why, and what your team should do this week. Every morning, on your own data.
01
The problem
A student doesn’t leave
overnight.
First, replies to messages slow down. Mock-exam scores drop, a payment runs late, attendance thins out. These signals are already recorded in your systems, but nowhere are they read together.
Noticing is left to a mentor’s intuition, and it usually comes too late. A student who leaves doesn’t cancel one payment; they erase the rest of the term.
02
Every morning
Three questions,
answered daily.
EO-Churn doesn’t send anyone a message. It ranks, gives the reason, and prepares the draft. The decision to send always stays with your team.
Why
The top three reasons for each student, by column: “message reply time +0.18”, “mock-exam trend +0.17”.
What to do
An outreach draft picked from your own approved templates, with a success measure for the conversation.
03
The dashboard
Everything in
one place.
Three views: the daily risk list, every student including those below the threshold, and model health: which threshold ran, how many rows were quarantined, when the last run happened.
04
How it works
From your export
to a mentor’s call.
No new data collection. EO-Churn reads what your systems already record.
- 01↘
Export
A CSV or LMS export from the system you already use is enough.
- 02↘
Column mapping
Your headers are mapped to the system’s schema once, with a single mapping file.
- 03↘
Training
The model learns from your own student history, not another institution’s. Unusable rows are set aside and reported to you.
- 04↘
Daily scoring
Every active student is re-scored automatically each morning.
- 05↘
Dashboard and e-mail
The ranked list, the reasons and the drafts land in front of your team.
- 06↘
Mentor action
The mentor reaches out and logs the outcome on the dashboard.
- 07↘
Feedback
The outcome feeds the next training run, so the model adapts to your institution over time.
05
The model
Chosen on your data,
not on ours.
Several candidate models are trained and measured on your data, and the one that performs best is selected.
01
Selection by PR-AUC
On imbalanced data, the real question is “who do I call first?”. PR-AUC answers it; ROC-AUC alone misleads.
02
Calibration
Raw scores are calibrated, so that “40%” really means 40%.
03
Cost-based threshold
A false alarm and a missed dropout don’t cost the same. The threshold is set from that cost ratio and cut to your team’s capacity.
04
Explainability
SHAP gives the top three reasons per student, by column name: “message reply time +0.18”.
06
Reliability
Principles we
don’t bend.
The rules that keep a backtest honest and a daily list usable.
No information
leakage
At prediction time the model can’t see any data that didn’t exist yet. Without this, backtests look misleadingly good.
A probability,
not a label
The system never just says “at risk”. It gives a calibrated percentage; your institution decides where to intervene.
Sized to your
capacity
If your team can hold twenty conversations a day, you get the twenty riskiest students. A list nobody can act on is no list at all.
Missing data
is reported
Missing data isn’t filled in. The record is set aside and flagged; if too much is missing, no list is produced that day.
Owned by your
institution
The model weights are handed over to you. The system keeps running even if our relationship ends.
Want to see it on your own data?
Request a free backtest ↗
07
Delivery
On their desk
by 09:00.
Your team doesn’t need to log in. The list arrives every morning on the channels you choose: all three, one, or none.
01E-mail
The daily risk list with per-student reasons, to several recipients. Your own SMTP server can be used.
02
Telegram
Instant notification to your team’s group or to one person. Your institution’s own bot, your own group.
03
Webhook
Slack, Discord or your institution’s internal system: any endpoint that accepts a message.
Your institution’s clock
Run time and time zone are configurable; the default is 09:00, Europe/Istanbul. Even on a UTC server, the mentor gets it at nine.
Weekdays only, if you like
It can run on weekdays only, so a quiet Monday morning doesn’t look like an outage.
We tell you if it goes quiet
If no successful run happens within a set window, a separate operator channel is alerted. A silent system and a healthy one look the same from outside; this check tells them apart.
WhatsApp is on the roadmap: it requires the Business API, sender verification and approved message templates.
08
Dashboard & API
Use our dashboard,
or don’t.
Every number on the dashboard comes from the system’s own API, and the same endpoints are open to you. They are protected with an API key and ship with an OpenAPI schema, so you can connect your CRM, student tracking system or reporting tool. We provide the dashboard; using it is up to you.
REST APIAPI KEY · OPENAPI
- /students
- Students above the threshold, with reasons
- /predict/{student_id}
- A single student’s live score
- /predict
- An instant prediction on the data you send
- /metrics
- The model’s performance metrics
- /health
- Whether the system is up
09
Outreach drafts
The list says who to call.
The draft, how to start.
There is no send button on the dashboard. The responsible team reads, approves or edits the draft; the decision to send is always human.
01
Open-weight model
Built on an open-weight model, not a closed service. It runs in your environment; no student data goes to a third party.
02
Adapted to you
It works with your approved communication templates and your terminology, so the output speaks in your institution’s voice.
03
It selects, it doesn’t write
The three SHAP reasons are matched against your approved template library. The model picks which template fits whom and prepares the draft.
10
Data & security
You decide how
data is shared.
We don’t ask for identity data; an anonymous student code is enough, and the key that links it to a student stays with you.
De-identified
sharing
Data is sent encrypted, without name, surname, e-mail or national ID number. Only behavioural data and the student code are shared.
Synthetic
data
You share no real records. Column names and summary statistics (mean, minimum, maximum) are enough; we generate the dataset.
Data never
leaves
The system is installed on your server or cloud account; training and prediction run in your environment.
- No column outside the mapping file ever reaches the model. You decide what you send.
- On request, raw data and everything derived from it is deleted after the backtest, with written confirmation.
- Once installed, data and model run in your environment; there is no outbound data flow.
11
Free backtest
NO FEE · NO SETUP · NO COMMITMENT
See it work on
your own data.
What we ask for
- Two CSVs: past student records and dropout dates
- Anonymized: no names, phone numbers or ID numbers
- One time only; no system installation
- A short form: fees, cost of outreach and team capacity
What you get, within 2 weeks
- How many dropouts we would have caught last term, and how many days early
- How long the daily list would be at your capacity
- A data-quality report: which columns are missing, how many rows are unusable
- If the result is weak, we’ll write that just as clearly
EO-Churn
All we need is
two CSV files.
Request a free backtest ↗
Student dropout early warning© 2026 EOAI Lab