AI-native learning, measured

AI learning programs with the evidence built in.

We run AI tutoring for schools and instrument every session, so institutions see measured outcomes, not vendor claims.

AssessmentsConversationsConfidence

Fig. 01 · The GlowingStar constellation

The problem

AI tutoring is everywhere. The evidence is not.

Learning that is measured, not assumed.

Fig. 02 · The calibration gap

Generic AI study tools are everywhere. Whether students actually learn, or just feel like they did, goes unmeasured. In our own classroom study, AI-tutored practice matched quiz-only practice on scores, while students grew more confident on the answers they got wrong.

Test-score gain

Quiz-only
+ AI tutor

Confidence when wrong

Quiz-only
+ AI tutor

Illustrative proportions; full figures in the evidence section

What we do about it

  • Tutoring enginemulti-agent plans, explanations, practice
  • Classroom deploymentteacher-controlled, runs as coursework
  • Pre/post assessmentwith per-question confidence ratings
  • Full telemetryevery message and attempt, logged

Product

One platform, indexed.

Fig. 02a · Session replay (reconstructed from real telemetry)

Student

Why does ice float on water?

Event log

  • 00:00.0session_started
  • 00:00.4message_sent

Demo transcript: a student asks why ice floats on water. The tutor explains hydrogen bonding and density, then serves a quiz. The student picks a wrong answer with high confidence. The system logs every event and flags the confidently-wrong response, the calibration signal a test score alone would miss.

Evidence

We report the inconvenient results too.

0

students in our first school deployment

0

subjects, taught in Traditional Chinese

0

student–tutor conversations

0k

interaction events logged

Test scoresboth groups improved similarly
Confidence when wronghigher in AI-tutored subjects
The gapexactly what our instrumentation catches

From a within-student comparison of quiz-only vs. quiz-plus-AI tutoring, run as real coursework. That honesty is the product: institutions should demand evidence from any AI learning tool they adopt, including ours.

Fig. 03 · Field study, Hong Kong, N=323

Builders × Researchers

The people who ship the tutoring engine also design the studies and publish the caveats.

Our team’s backgrounds span Harvard, MIT, Stanford, and the University of Toronto, including work through the MIT Media Lab and the Harvard Innovation Labs.

HarvardMITTorontoStanfordMIT Media LabiLab

“Every student is a glowing star. Our job is to help them shine.”

Our founding conviction

Fig. 04 · The workshop

Contact

Bring measured learning to your institution.

Onboarding a small number of partner schools and universities for upcoming terms.