Answers tied to a student's own enrolment

Answers and access both follow the enrolment record

In a straightforward retrieval design the question alone decides the answer. In education the asker decides it too: when an assessment is due, how many credits are needed to progress and whether a resit is permitted differ by programme, cohort year and delivering centre. The enrolment record therefore belongs inside retrieval: the asker is resolved to their enrolments, then the corpus is filtered to what is in force for that cohort and module instance.

The same enrolment data is the access filter, applied before scoring rather than over results. A passage the filter admits is a passage the model may quote, so the boundary lives in the filter. Staff are not one tier either: a module leader sees their own module's marking scheme, a tutor at a partner centre only their delivery.

The system answers documented process questions. It does not mark student work, decide an extenuating circumstances claim, an appeal, a fitness-to-study case or an academic misconduct case, or give immigration advice. A disclosure of distress or self-harm, or a safeguarding concern about a minor, is routed to your named wellbeing or safeguarding contact rather than answered.

Handbook versions and the regulations a cohort sits under

A continuing student is normally assessed under the regulations in force when their cohort started, so recency alone cannot decide which handbook answers a question. Validity against the cohort decides it, and the newest version is right only for the cohorts it governs.

Each document therefore carries a validity window of academic year, cohort and programme, recorded against the module instance because codes are reused across years. Superseded versions stay retrievable for the cohorts they still govern, and the citation names the module and version.

Student records, the learning platform and the standards between them

The data usually sits in at least two systems whose records can diverge: the student information system holds enrolment, fee status and progression, the learning platform content, groups and the gradebook. Which is authoritative is decided field by field.

LTI 1.3 launches a tool inside a course with a signed role and context claim. Where LTI Advantage is enabled, its names-and-roles service supplies membership and its assignment-and-grade service can return a result; LTI 1.1 has neither, so what a deployment exposes is checked.

Enrolment itself usually comes from the student information system, by API, middleware or scheduled extract; OneRoster is the common route where the provider is a school. Because enrolment changes through add and drop, transfers, interruptions and resits, permissions re-synchronize on that feed, on a cadence agreed separately from content refresh, and a nightly extract caps how fresh the filter can be, so that window is agreed in the proposal.

Admissions and enquiry load at the seasonal peaks

Enquiry volume concentrates into results and late-admissions periods, enrolment week, the opening of term and assessment deadlines rather than across the year, and many of those enquiries carry a deadline. Sizing therefore uses your own queue history for those weeks, with queue depth and the escalation path set against the peak day.

An applicant is a different case from an enrolled student: no enrolment record, possible duplicate applications, and an offer conditional on results not yet published. Identity resolution therefore runs against the admissions system, and conditional offers route to a person.

Where administration ends and assessment begins

Administrative questions have a documented answer: a deadline, an accepted format, how to submit an extenuating circumstances claim, a credit requirement for progression. The decisions that follow do not. Assessment is academic judgement, subject to moderation and external examiner scrutiny, and the outcome of a claim, an appeal or a misconduct case belongs to the panel, so the boundary is enforced in code.

On the assessment side, what can be built is support for the marker: draft feedback referenced to the rubric, or a flag where a submission appears to miss a stated requirement. No mark is generated, and a write back through an assignment-and-grade service needs approval.

Student data, consent and retention periods

Student records are personal data about identifiable people, and where the provider teaches under-18s they include data about minors. Which regime applies depends on where the provider and students are, and they are not parallel: FERPA governs disclosure of US education records; UK or EU GDPR requires a lawful basis, sets a children's consent age and limits transfers; the DPDP Act requires verifiable parental consent for under-18s. A US provider teaching school-age learners is also in scope of COPPA and state student-privacy laws. We scope these with you rather than advise on them as a legal matter.

Retention is asymmetric: the award record and its assessment evidence are kept for long statutory or awarding body periods, conversation logs usually should not be. The proposal also records whether student submissions may reach a model provider.

Evaluating against the academic year

The evaluation set comes from questions student services and admissions already answer, not invented ones. It carries the cases specific to this sector: one question asked by two students on different cohorts where the answers differ, and a question answered only by a superseded handbook. It also carries the access cases, scored as a refusal plus a route to a person rather than a retrieval miss: a student asking for another programme's assessment brief, or a tutor asking for a marking scheme outside their own delivery.

Rollover is a regression event. Codes change, new handbook versions are approved, cohorts advance and last year's modules are archived. Running the same set afterwards shows whether validity filtering and the access filter still hold.

What a project for an education provider costs

Every engagement starts with a fixed-price discovery sprint at USD 2,500 over two weeks, which produces the architecture, initial screen designs, acceptance criteria, a running-cost model and a fixed price for the build. The sprint fee is deducted from the build.

Answering cost follows enquiry volume, which arrives in concentrated weeks rather than spread evenly. Indexing cost recurs at each academic year rollover, when much of the corpus is replaced, and again where partner centres maintain separate handbook versions.

Part of the AI Feature Implementation Sprint → · Explore our work → · Discuss an AI project →