matriculate
equipping near-peer college advisers with frontier AI that is emotionally sensitive, and accountable.
Matriculate trains undergraduate volunteers (Advising Fellows) to help low-income high-school students get into and pay for university. The work is intimate, asynchronous, and run almost entirely over text. Our engagement is the careful build of the AI layer that sits behind those Fellows: a system designed to enhance their judgement, not replace it, and held to the standards of the advisers themselves.
Two things make this work delicate. Advising Fellows are themselves recent first-generation university students; the closeness of their experience to the student's is the asset. And the moments that matter most (disclosures of crisis, financial-aid panic, family pressure) are exactly the moments where a fluent but emotionally flat model would do the most damage. The standard adoption pattern of "wire up a chatbot, learn from the failures" is not available here.
The engagement delivers four canonical artefacts: a response-evaluation rubric, an advising-fellow survey synthesis, a corpus of five hundred gold-standard responses, and the project management to ship it. Three supplementary artefacts have been absorbed into the same envelope.
The rubric, twelve dimensions across two registers, anchored at five points, with two hard floors, is built directly on seventy Advising Fellow survey responses and the Matriculate AF Code. Every anchor point is sourced; provenance markers distinguish verbatim AF voice from adapted-descriptive and synthesised passages. A crisis & distress response playbook (twelve sections, AF voice grounded in survey material) and a college-process accuracy answer key (ten sections, financial-aid dominant per AF priority) sit alongside.
The five hundred gold-standard responses are the training corpus the chatbot will be fine-tuned against. Each is hand-scored against the rubric and authored to model the register Matriculate wants its AI to inhabit.
The most useful early finding came from the AF Survey itself: across seventy responses, "break the big thing into the manageable thing" appeared as a near-universal AF behaviour, and was almost entirely absent from the existing chatbot prompt. A single behavioural pattern that the AFs do constantly and the model did not do at all. The rubric now anchors it as its own dimension.
Phases 1 and 2 complete; Phase 3 (gold-standard authoring) underway. Production launch targeted for 1 July 2026, after which the rubric becomes the evaluator dashboard's spine.