We publish the measurements, including the bad ones.
Five areas, six full papers, and a standing commitment to say when something did not work. Two of our papers report a negative result, one of them about a model we had already deployed.
Five areas, and what we don’t know yet.
Each area has a lead, a written agenda and a list of problems we have not solved. We publish the open problems because a research group that only lists its answers is a marketing department.
Predictive modelling
Fitting persistence, completion and course-risk models to a single institution, validating them against a baseline that does not require machine learning, and knowing when not to ship.
Predictive modelling
Open problems we are working on
- Warm-starting for institutions with under two terms of usable history
- Calibration that survives a mid-year change in academic regulations
- Honest uncertainty communication to an advisor who has 400 students and four minutes
Selected work
Causal inference & efficacy
Measuring whether an initiative actually changed anything, at a scale and cadence an institution can sustain, using designs an accreditor will accept.
Causal inference & efficacy
Open problems we are working on
- Efficacy designs for universally-offered interventions with no comparison group
- Detecting effect decay early enough to act within the same academic year
- Pooling underpowered small-programme studies without hiding heterogeneity
Selected work
Fairness & measurement
Auditing who a model serves worst, closing the gap, and being explicit about the attributes we cannot audit because nobody records them.
Fairness & measurement
Open problems we are working on
- Fairness auditing when the protected attribute is unrecorded and proxies would harm
- A gap definition that does not treat historical inequity as ground truth
- Caste, disability and language dimensions in the Indian context
Selected work
Curriculum & capacity
Finding the structural obstacles to on-time completion — full sections, rigid prerequisite chains, courses offered once a year — and forecasting demand well enough to fix them.
Curriculum & capacity
Open problems we are working on
- Estimating suppressed demand from students who never attempted to register
- Multi-year timetable optimisation under staffing constraints
- Distinguishing a hard course from a badly scheduled one
Selected work
Language & student voice
Making sense of free-text feedback at scale, in the multilingual and code-mixed reality of Indian classrooms, without producing per-student scores we would not defend.
Language & student voice
Open problems we are working on
- Topic modelling that is stable across code-mixed English and Indian languages
- Reaching students who left and therefore never filled in the form
- Aggregate insight that cannot be re-identified to an individual
Selected work
Full papers, most recent first.
Every paper carries an abstract, a stated method, figures with the underlying data, an explicit limitations section and its artefacts. See the full record →
Method first, headline second.
We sell analytics, so we have a commercial interest in these results. Pretending otherwise would be worse than admitting it. These are the rules we hold ourselves to so the work stays checkable by an institutional researcher who does not trust us.
The outcome is fixed before analysis
For every efficacy study, the outcome measure and design are recorded before we look at the data, so a disappointing result cannot quietly become a different question.
We cannot share student records, so we share a stand-in
Real institutional data never leaves its tenant. Where a result cannot be reproduced without data, we publish a simulated corpus that reproduces the qualitative finding.
Every paper says where it is weak
Self-selected samples, Indian skew, matched comparison that is not randomisation. If a caveat would change how you read the number, it belongs in the paper.
Papers are versioned in public
When a reader finds an error we revise, bump the version and say what changed at the top. Four of the six have been corrected this way.
Who does the work.
Four researchers and three external advisors. Small enough that every paper has a person you can write to about it. Read more →
Disagree with a number? Tell us.
We would rather be corrected in public than wrong in private. Every paper lists a correspondence address and a person who will read it.
Replication data and harnesses available on request