Everyone looks
equally present.
A large online lecture. Every camera a black tile. The instructor is talking to a wall he can’t read, and can’t tell who’s actually still with him.
Attendance is visible. Learning isn’t.
In live online lectures of 100+ students, instructors have no reliable way to know when students feel unseen, confused, passive, or disengaged. It stays invisible until it shows up as a poor exam or a dropped course, too late to act.
The problem was never content quality. It’s that nobody notices the moment a student checks out.
Meet the instructor living that gap.
- , He needs to know, in real time, who's confused or checked out.
- , By the time a quiz shows him who was lost, the session is already over.
- , Every time he tries to re-engage the room, it's a guess, he can't tell if it worked.
- , Cold-calling is his only live signal, but it breaks his teaching flow.
So what would it take to close that gap?
Restore the ability to read a room the moment class goes online. Turn attendance from a task-driven obligation into a momentum-driven learning experience, by giving instructors the awareness a physical classroom gives them for free. The adaptive layer comes later, and only after the diagnostic has earned it.
A co-pilot for the classroom, answering his pain directly.
A real-time picture of the class in three states, so the instructor knows where to direct attention right now.
Suggests the next best action when a group drifts: a poll, a break, a check-in. The instructor stays the pilot.
Zoom API and Brightspace integration. No new platform to learn; AHA lives inside the class he already runs.
On-device processing, disclosure at session start, session-level deletion, and the Two-Witness Rule.
AHA is an attention intelligence layer on the live class.
Real-time engagement per student, an aggregate class score, and a one-click private ping.
When the class dips, it suggests a recovery move: a poll, a cold-call, a recap.
Detects eyes-away, drowsiness, and confusion. All processed locally, never transmitted.
What the instructor sees.
The working prototype at aha-moment-nine.vercel.app, the same loop, running in the browser right now.
Isn’t this just surveillance? We only see engagement, not students.
Every intervention is one click and instructor-approved. The Two-Witness Rule means a facial signal alone never fires an alert.
On-device processing. Zero raw video stored or transmitted. Only the engagement state ever reaches the instructor.
Per-student baselines tested for disparities across disability, neurodivergence, lighting, and skin tone; consent tracking, IRB & FERPA readiness.
Torres, again. What does his class look like now?
- , Now he knows, in real time, who's confused or checked out.
- , The session isn't over before he finds out who was lost.
- , He's no longer guessing, he knows if the intervention worked.
- , Cold-calling isn't his only signal anymore, and his flow stays intact.
One North Star, three KPIs, hard guardrails.
The delta between AHA-on and AHA-off sessions, normalized per student and per session so a 200-student lecture and a 40-student seminar compare fairly.
Share of alerts instructors judge correct. Validated in Pilot against Live Student Pulse as ground truth.
Share of alerts that lead to a real action. An unused alert is a failed alert.
Repeat weekly use. Engagement minutes mean nothing if instructors stop showing up.
Instructors create demand. Institutions sign contracts.
Free semester-long design-partner pilots at 3 universities via warm academic channels. Advocacy before any sale.
The Zoom API integration puts AHA inside the tool instructors already open every day. Friction drops; visibility compounds.
EDUCAUSE-type IT & procurement conferences reach the legal, privacy, and IT approvers instructor enthusiasm can't.
From black tiles to the AHA moment.
Where AHA wins, and what we refuse to copy.
Key differentiators: the Two-Witness Rule (a delighter no single-signal competitor offers), privacy-first on-device processing, a cameras-off behavioral fallback, and alerts designed for instructor action rather than student punishment.
Three twists that shaped the product.
Our learner-side completion tool failed Cagan's value risk: people didn't want it badly enough.
Pivoted to the instructor side: detect the moment of disengagement instead of fixing content.
Redefined the problem, the persona, and the entire MVP.
Team changed after the topic and early interviews were set; new members lacked discovery context.
Split ownership by strength and documented decisions so context transferred fast.
The competitive analysis and buyer-side thinking exist because of the reshuffle.
A senior stakeholder proposed Live Student Pulse: students self-report their own state.
Accepted, but scoped as a third independent validation signal, not a trigger.
Pulse becomes Pilot-phase ground truth for Flag Precision Rate, without letting one click override two agreeing signals.
The commercial detail behind the one-liner.
We never recalibrate off one mismatch.
Live Pulse is a third, independent signal, it validates the Two-Witness engine but never triggers an alert on its own. So we apply Two-Witness logic to the fairness check itself: one signal can never override another alone.
Self-report and quiz results agree the student is fine, so a Drifting flag is likely the miscalibration. We treat their baseline as a recalibration candidate, but only if the pattern repeats across sessions, never off one class (which could just be prior knowledge).
Not the AI being wrong, not the student lying, engagement ≠ learning, the exact gap we named. The click can't be trusted as a learning signal, so it never recalibrates anything.
Recalibrate a baseline only when the same mismatch repeats across sessions and a third signal (quiz / comprehension) agrees. One signal never overrides two.
Human-in-the-loop, where the human is the student too.
Closes the accessibility gap, and gives a faster second signal alongside quiz data to check detection.