Three layers that only work together.
A warehouse with one agreed definition of every term. Models fitted to your institution rather than a sector average. And workflows that turn each insight into an outreach somebody owns. Remove any one layer and the other two stop mattering.
What each layer actually does.
Each one is documented in full on the feature reference, including the limits we know about.
Models fitted to your students.
Predicted persistence and completion for every enrolled student, refreshed nightly, with the contributing factors named so an advisor can argue with the number rather than obey it.
Insight that someone owns.
Each flagged student becomes an outreach with an assignee, a channel, a due date and a recorded outcome. Shared case notes stop three departments contacting the same student in one week.
One warehouse, one dictionary.
SIS, LMS, CRM, attendance, assessment, fees and card data consolidated into a cloud warehouse with one versioned definition of every term. This is the step most projects underestimate.
Fitted to your students, not a benchmark.
A model trained on one institution loses about 19 points of precision when moved to another — we measured it across 14 contexts and published the result. That is why every deployment gets its own fitted model, validated on your held-out data before anything reaches an advisor.
Persistence, completion, and course risk
Three outcomes rather than one composite risk score, because a student at risk of dropping a course is a different problem from one at risk of leaving.
Contributing factors, named
Every prediction carries its top factors. An advisor sees “two missed assessments in a gateway course”, not “risk 0.71”.
Beaten against a baseline first
If the fitted model does not beat a simple attendance-and-assessment baseline on held-out data, we say so rather than shipping it.
Checked by subgroup, every refit
Flag rates and precision are reported by first-generation status, gender, programme and campus. Our own first model under-flagged first-generation students by 11 points.
An insight nobody owns changes nothing.
The most common failure we see is not a bad model. It is a good dashboard that nobody acts on, because acting on it was never anybody’s job. Workflows exist to close that gap and to leave behind clean data for the next efficacy study.
Prioritised by predicted impact
Not alphabetically, and not by raw risk — by where an intervention is most likely to change the outcome.
Assignee, channel, due date
Every outreach has a person attached. Unclaimed items escalate rather than expire quietly.
Shared notes across departments
Advising, faculty and student services see the same history, so a student is not contacted three times in one week by people who do not know about each other.
Outcomes recorded, not assumed
What was tried and what happened. Without this the next efficacy study has nothing clean to work from.
The step everyone underestimates.
Most failed analytics projects fail at data, not at modelling. Our first deliverable is usually not a model — it is a data dictionary that your registrar, deans and IR office all agree to sign. Nothing downstream is trustworthy until that exists.
What the first ten weeks look like.
Honest timeline for a single-campus institution with reasonable data hygiene. Multi-campus systems with heterogeneous SIS instances take three to five months, and we say so in the proposal.
| Phase | Weeks | What we do | What you do |
|---|---|---|---|
| Discovery | 1–2 | System inventory, data access, definition workshop | Name one owner per source system |
| Integration | 2–5 | Connectors, warehouse build, lineage | Approve the data dictionary |
| Validation | 5–7 | Reconcile against your own reporting; fix what disagrees | Sign off that the numbers match yours |
| Model fit | 7–9 | Fit, validate against baseline, subgroup fairness check | Review the factors for face validity |
| Go live | 9–10 | Dashboards, queues, advisor training | Decide who owns outreach |
| First efficacy study | +1 term | Matched-comparison assessment, published to you in full | Accept the result, including a null one |
See it against your own data.
We will run a short discovery on one programme and show you what the platform would surface. If your data is not ready, that is what we will tell you.