Learning recovery, as institutional infrastructure.
Why now
Four shifts, one gap.
Grounding is becoming standard.
LMS vendors are shipping assistants that answer from course material. The step after the explanation — retry, recovery, hand-off — has no owner.
Accessibility became a buying gate.
The ADA Title II rule sets WCAG 2.1 AA for the web content public colleges provide, including through vendors, with deadlines from April 2027.
Completion is funded and measured.
Title III and V, state student-success awards and Workforce Pell programs all reward interventions an institution can evaluate.
Support is staffed for office hours.
Tutoring centres and SI programs reach the students who show up. Students get stuck at midnight, the week before the exam.
Where it sits
Between the LMS and the people.
The LMS owns the course shell. Publishers own content. Assistants explain. Student-success teams own the phone call. Zelvi is the layer that turns a wrong answer into a recorded recovery or a hand-off with the evidence attached, and it launches from the LMS rather than competing with it.
The asset compounds. Every loop writes a structured record: misconception, intervention, retry, result. Across courses and institutions, that becomes evidence of which explanation resolves which misconception.
- Student success & tutoringAdvisors, SI leaders and tutors decide who to call.
- Zelvi — the recovery layerMisconception → re-teach → retry → recorded result → hand-off with context.
- AI assistantsExplain, summarise and generate practice.
- Course materialYour syllabus, OER, publisher content, SOPs.
- LMSCourse shell, roster and gradebook.
What is built
A working product, not a deck.
- 100% vs 54.2%
- answers that stayed inside the course material on our benchmark, against the same model handed the whole document.
- 54 + 27
- STEM diagrams and interactive figures drawn live from the concept, not pasted images.
- LTI · SSO · SIS
- LMS launch and grade passback, SAML single sign-on, OneRoster, Clever, ClassLink and Google Classroom.
- 2 misses
- before the model is called. Rules and authored hints go first, so AI cost tracks struggle, not chat volume.
- 3 languages
- of crisis-language screening on every typed and spoken turn. Ava stops, says she is an AI, and alerts staff.
- 22
- languages Ava replies in, with the course’s English terms kept beside the translation.
Go to market
Land one course, expand by evidence.
- Step 1Design partnerOne course, one term, no cost
- Step 2Paid courseThe same course, the next term
- Step 3ProgramEvery section of the gateway sequence
- Step 4InstitutionAcross programs and campuses
- Step 5SystemState systems and consortia, one protocol
Community-college gateway STEM
College Algebra, gateway biology and chemistry: high enrolment, a tracked pass rate, and a retry that can be scored objectively.
Allied-health prerequisites
Anatomy, pharmacology and terminology, where a failed prerequisite ends a career path before it starts.
Short workforce programs
Accelerated healthcare and technical credentials, where there is no time to recover from a bad week.
Where we are
Product live. Pilots next.
The loop runs end to end on College Algebra and allied-health foundations. We are recruiting the first design-partner institutions now. There is no efficacy data yet; the first pilots exist to produce it, measured with the institution’s own research office.
The next milestones
- 1Two to three institutional design-partner pilots
- 2The first paid conversion
- 3A gateway-course recovery study run with an institutional research office
- 4The accessibility report reviewed by an institution’s accessibility office