How a curriculum-grounded AI tutor answered 70–80% of student questions, safely

Every edtech board wants an AI tutor. The blocker is rarely the model, it's the risk. In education, a confidently wrong answer mis-teaches at scale, and the users are often minors whose data is protected by law. This provider needed AI that personalizes learning without hallucinating content or exposing student data.

The Challenge

Consumer AI tutors were already in students' hands, ungoverned, off-curriculum and outside the platform. The provider had to offer something better and safer, and three problems stood in the way:

1. 01

   Hallucinated teaching content. A general-purpose model would confidently teach facts and methods that didn't match, or contradicted, the approved curriculum.

2. 02

   Minors' data and compliance. Any tutor touching student data carried FERPA and COPPA obligations, and sending that data to a public model was a non-starter.

3. 03

   Safety and accountability. Responses had to be age-appropriate, and teachers needed to stay accountable for grades, feedback and anything escalated.

They didn't want a chatbot bolted onto the LMS, they wanted a tutor that only teaches what the curriculum says, keeps student data inside the walls, and always has a teacher in the loop. The engagement brief

The Solution

Periscope built a retrieval-grounded AI tutor: every answer is drawn from the provider's approved curriculum, wrapped in age-appropriate guardrails, evaluated for accuracy, and deployed inside the provider's boundary, with educators reviewing feedback and every escalation.

Retrieval-grounded answers. The tutor reasons over the approved curriculum and cites it, so responses come from validated material instead of the model's imagination.

Age-appropriate guardrails. Safety filters, content boundaries and red-teaming keep every response suitable for the learner in front of it.

Evaluation & drift monitoring. Eval suites score accuracy against curriculum answer keys, and drift is caught before it reaches a student.

FERPA / COPPA data handling. Private deployment, minimization, consent and audit logging keep minors' data inside the boundary and every access accountable.

Educator-in-the-loop. The tutor proposes explanations, feedback and flags; teachers approve grades and handle escalations, a human stays accountable for what reaches a child.

The Results

Grounding the tutor in the curriculum, and keeping teachers in the loop, turned a risky idea into a safe, always-on learning aid students and parents could trust.

Instant help, on-curriculum. The tutor answers 70–80% of routine student questions instantly, drawn from approved material.

Hallucinations contained. Retrieval grounding and evaluation kept off-curriculum and incorrect answers from reaching learners.

Compliant by design. Student data stayed inside the provider's boundary, with the audit trail FERPA and COPPA reviews expect.

Teachers freed, not replaced. Educators moved from answering the same questions to the high-value work, with the tutor handling the routine and routing the rest to them.

A curriculum-grounded tutor, guardrailed and inside the FERPA/COPPA boundary, answering 70–80% of questions instantly, without hallucinating content or exposing a single student's data. Outcome summary