Built for the institution, earned with the learner.
Every incumbent in medical education was built for an individual learner with a credit card. None was built for the institution that has to teach, staff, accredit and re-license clinicians for decades. That gap is the whole reason this exists.
Three broken workflows, one root cause
A student learns on one bank. A resident learns on another. A physician chases CME on a third. Nothing carries forward. Four years of calibrated performance data becomes a line on a transcript, and the residency program starts from zero — as does the learner.
Meanwhile the program director rebuilds milestone evidence by hand every cycle, the registrar reconciles rosters by CSV, and the faculty member who wants to write questions about what they actually taught last week faces a blank page and a four-hour job. The learning is continuous. Only the record is broken.
So VeloLibrary starts from the record rather than from the question bank. One learner, one longitudinal record, and every surface — practice, cases, review, cohort analytics, milestone evidence, CME — reading and writing the same rows. The accreditation output an institution needs stops being a reporting project and becomes a by-product of ordinary learning.
How we build
Principles that are enforced, not aspired to
Each of these corresponds to a control in the codebase — a constraint, a trigger, a test that blocks the deploy. A principle nobody can point at is a slogan.
The record is the product
Every response, review, case turn and credit lands in one longitudinal learner record. Features write to the record; portals read it. That is what makes school-to-residency-to-CME a continuity rather than three purchases.
Explain or do not ship
Every adaptive decision carries a human-readable reason. Adaptivity that cannot explain itself is indistinguishable from a shuffle, and a faculty member cannot defend it in a grade appeal.
AI drafts, humans sign
AI never publishes an item, issues a credit, or files a milestone rating on its own. Every generative output enters a review state owned by a named person — enforced in the schema, not in a policy document.
Respect clinical time
The learner is often post-call. Sessions resume mid-block, practice works offline on hospital Wi-Fi, and notifications honour duty-hour quiet windows. A ten-minute session has to be a complete, useful unit.
Psychometrics or it did not happen
Item quality is measured nightly — p-value, discrimination, distractor analysis — and flagged items reach faculty automatically. Prediction is calibrated and reports its own error rather than a percentile.
Sovereignty is a feature
One codebase, two modes. An institution that cannot let learner records leave its infrastructure runs the same product inside its own Azure tenant, not a reduced version of it.
Definition of done
What we are building toward
V1 is done when a pilot medical school runs a semester on VeloLibrary: learners practise daily from an AI-generated queue against a calibrated adaptive bank with spaced repetition; faculty author items with the AI assistant and watch cohort heatmaps; the registrar syncs the roster from their SIS nightly; and the program director exports a milestone evidence packet for a CCC meeting — all inside their own tenant, themed to their brand, with every action in the audit log.
Where we are. The learner loop, the authoring engine, the case tutor and the AI layer are built and running in the demo tenant. We are piloting with medical schools for the coming semester, which is why you will find targets rather than customer numbers on this site.
Part of the Velo family
VeloLibrary shares its design system, tenancy model and AI gateway with VeloDesk, the healthcare IT service platform. Same engineering standards, same isolation guarantees, different problem.
Who we build for
Medical schools, residency programs and health-system CME offices — the three buyers nobody currently serves together. The institution is the customer, and the product still has to earn its place with the learner every single day.
What we will not do
We will not let a model publish content, rate a milestone or issue a credit. We will not sell or train on learner data. And we will not ship an adaptive decision we cannot explain to the person it was made about.