Case study

AI tutor for Online School - guides, never gives the answer

An AI tutor that walks the student towards the answer, tracks what they keep coming back to, and reports to the teacher who needs help with what. Memory carries across grades - the system that knew a student in 4th grade is the same system supporting them in 7th.

Live · AI Tutor

900+

students across grades 4-7

Quiz-score lift

+27%

vs pre-system baseline

Per teacher, weekly

10-15h

reclaimed

Last updated:

01. The problem

Three breakdowns the off-the-shelf tools couldn't fix

The school provided fully online classes. They were offering a combination of pre-recorded videos, a generic chatbot, well-crafted materials and curriculum. Three problems with that combination:

  1. Generic chatbots give the answer. Every off-the-shelf assistant the team tried would happily provide the answer rather than guide the student through the problem and point him to the right location where they can find the answer. Students learned how to prompt; they didn't learn the material.
  2. Teachers had no insight. A weekly quiz gives a snapshot of who passed and who didn't. It doesn't tell the teacher why a specific student keeps stumbling, or which three students would benefit most from a 30-minute one-on-one on a specific topic.
  3. Knowledge of the student didn't carry forward. A 4th-grader's strengths and weaknesses live in one teacher's head. By 7th grade - several teachers and several class transitions later - that history is lost.

Build it under FERPA, COPPA, and the relevant state-level under-18 data laws, and lay the encryption-at-rest and access-control architecture before any student record is written.

The Brief

Build a tutor that respects the student's effort, gives the teacher real signal, and remembers the student across the four years.

02. The stack we built

Coach for the student, signal for the teacher, memory across grades

  • The tutor agent

    A frontier reasoning model for one-on-one coaching; smaller models for routine quiz generation. The architecture is model-agnostic and the system has been continuously updated as newer model families have landed. The agent has access to the curriculum, the student's history, and a strict instruction layer that forbids handing out the answer. It rephrases, hints, asks the student to try a step, and only escalates to a worked example when a teacher-set threshold is met.

  • Per-student memory

    Persistent record of which topics the student has revisited, how often, what mistakes they made, what they got right on the second try, what they keep coming back to weeks later. This memory is the system's understanding of the student; the tutor reads it before every session.

  • Dynamic quiz generation

    Quizzes are not fixed. Each student gets a quiz tuned to their current weak spots.

  • Teacher dashboard

    The teacher sees the class as a heat map: who's stalled where, who's racing ahead, which two or three students would benefit from intervention this week, which topics the class as a whole is weakest on.

  • Reports that learn the teacher's voice

    The system tracks which report formats each teacher prefers (which sections they keep, which they delete) and produces the next report in that shape.

  • Cross-grade carry-over

    When a student moves from 4th to 5th, from 5th to 6th, from 6th to 7th - their tutor history moves with them. The new grade's teachers inherit a written summary of the student's strengths, weaknesses, and learning preferences.

  • LMS integrations

    The school operates a custom LMS platform on top of Canvas. The tutor reads from and writes to both, normalising assignments, gradebooks, and student records into a single internal schema.

03. Compliance posture

Compliance as architecture, not as a filter

US under-18 student data is one of the most regulated categories of data on the public internet. The compliance posture is the system, not a footnote.

FERPA-compliant data flow

Records are tied to the school as the controller; CloudSource is a school-official processor under written agreement. No advertising signals, no cross-tenant data sharing, no model-training on student records.

COPPA-compliant consent flow

Parental consent collected per child before any record is written; consent state is a first-class field on every student row.

State-level under-18 laws

Several of the school's enrolled students live in states with stricter-than-federal data laws. The system enforces the strictest applicable jurisdiction per student - not a global lowest-common-denominator.

Encryption at rest

All student records, all chat transcripts, all quiz histories - encrypted at the storage layer with per-tenant keys. Key rotation on schedule.

Encryption in transit

TLS-only on every API hop, including internal service-to-service calls.

Audit log

Every read, every write, every model-call against student data is logged with the calling user, the time, the purpose, and the data fingerprint. Six-year retention to match the school's own record-keeping policy.

04. Implementation

Three months, four phases - each ended with a working layer

  1. Shipped
    01Weeks 1-4 · Tutor

    Built the core tutoring agent. Curriculum ingestion across the four grades, lesson-aware context, and dynamic quiz generation tied to each student's current weak spots. The hint-and-coach dialog flow with a strict instruction layer that forbids handing out the answer: the agent rephrases, hints, asks the student to try the next step, and only escalates to a worked example when a teacher-set threshold is met. End of phase: a tutor that handles a real lesson end-to-end without a person in the loop.

  2. Shipped
    02Weeks 4-7 · Memory

    Wired the memory model - short-term per-session and long-term per-student. The tutor reads the student's full learning history before every session: which topics they've revisited, how often, what mistakes they made, what they got right on the second try, what they keep coming back to weeks later. Cross-grade carry-over: when a student moves from 4th to 5th, from 5th to 6th, from 6th to 7th, their tutor history moves with them. The new grade's teachers inherit a written summary of strengths, weaknesses, and learning preferences. End of phase: a tutor that can follow a student across the four years of the curriculum.

  3. Shipped
    03Weeks 7-10 · Compliance

    FERPA, COPPA, and state-level under-18 data laws baked into the architecture, not bolted on. Per-tenant encryption at rest, TLS-only in transit, audit log on every read, write, and model call against student data. Parental consent as a first-class data store; the system enforces the strictest applicable jurisdiction per student rather than a global lowest-common-denominator. The pipeline branches at the consent boundary so PII never leaves memory unless a consent record is on file. End of phase: a system the school's lawyers signed off without asterisks.

  4. Shipped
    04Weeks 10-12 · Teacher monitoring

    The teacher dashboard. Class as a heat map - who's stalled where, who's racing ahead, which two or three students would benefit from intervention this week, which topics the class as a whole is weakest on. Plus a chat agent the teacher queries directly: "which students need help with fractions?" / "what's <Student> been stumbling on the past two weeks?" - answered against the actual memory and quiz history, with citations to the underlying sessions. Reports learn each teacher's voice over time (which sections they keep, which they delete) and produce the next report in that shape. End of phase: full handover to the teaching staff.

3 months from kickoff to ship. Live and continuously updated since early 2024.

05. The numbers

Real classroom traffic, real teacher time saved

Project length (initial build)
3 months
Students currently tutored
900+ across many classes
Teachers using the dashboard daily
30+
Quiz-score improvement vs baseline
+27%
Teacher-time reclaimed per week
10-15 hours per teacher
Compliance posture
FERPA · COPPA · state-level under-18 laws · encryption at rest · audit log
AI that already runs

A different shape of student-facing system?

The tutor pattern - strict instruction layer, persistent per-user memory, role-aware dashboards, hard compliance posture - is the same architecture we ship for customer-facing internal copilots and complex multi-stakeholder workflows where the agent must guide rather than answer. A 30-minute diagnostic walks through your specific shape and tells you whether this fits.

Live · powered by Gemma-4 · running on our hardware in Sofia