Ebb

Inspiration

For five years I worked as a tutor, and over that time I worked with many students with special needs. I also got to know their parents, several of whom worked as instructional assistants (IAs) in schools. Those conversations stayed with me. They described a community that is deeply committed but often stretched thin: students who get overwhelmed in the middle of a lesson and can’t always find the words to say so, and adults who are trying to notice and respond to many needs at once. They also described how hard it can be to keep strategies consistent between home and school.

What stood out most was that none of this came from a lack of care. The families and educators I spoke with cared enormously. What was missing were tools that fit the moment, ones that are personal to each student, discreet, and respectful of their dignity. I kept thinking about how a student might get a quiet, private way to reset, and how an educator might get a gentle way to check in without singling anyone out.

That is why we built ebb. It is a wearable and companion app that lets parents and teachers set up a student’s calming preferences, send supportive prompts, and give students a way to ask for help on their own terms. We chose techniques with real support behind them, like slow breathing and predictable routines, and we made privacy and sensory safety central to the design.

I wanted to take action instead of only hearing about the problem. Ebb is our first step toward a solution that is built with the community’s voice at its center, and we hope to keep improving it alongside the families and educators who inspired it.

What it does

Ebb is a smartwatch, an iPhone app and a website that help kids with special needs calm down before a meltdown, and keep their parents and teachers informed.

On the watch

Senses rising stress: it reads movement (rocking, hand-flapping) and room loudness, and compares them with what’s normal for this child. It keeps loudness levels only, never audio. Responds in steps: a gentle breathing buzz → a full calm mode (a breathing circle, a calm scene, a soothing voice, light vibration) → a “needs an adult” alert if it doesn’t settle. Starts calm three ways: by itself, when the child taps the watch, or when a teacher or parent taps Start in the app. Asks the child afterwards: three faces for better, same or worse.

It learns

Measures results: how much agitation dropped and how fast the child settled. Tunes itself: “False alarm” or “Helped” ratings from teachers make it less or more sensitive. Picks scenes: it learns which scene helps this child most at each time of day, and still tries new ones sometimes. Predicts: it forecasts the riskiest hours ahead.

For the adults

Parent setup: pair the watch with a QR code; fill in an assessment, and AI drafts a calm profile; choose the calm voice (an AI voice or their own cloned voice); approve 6 scenes each week. Teacher tools: start or stop calm, rate each moment, write journal notes. Live view and insights: a live agitation graph, a heatmap of when hard moments happen, what works, and an AI weekly report. Texting: parents and teachers can text Ebb (“How was Sam today?”) and pass notes to each other by iMessage. Location (optional): GPS only if the parent opts in; teachers see it only during school hours.

Privacy: the child is known by a nickname only, there’s no audio, the AI never receives names, and data is automatically deleted after a set time.

How we built it

We got a programmable watch, created a ui for the website, app and watch. Connected the backend and the logins. We used photon for the messaging system. We used gemini for the profile for the assessment for the student done by the parent. We used the Arduino ide to push all the updates to the watch. We also used elevenlabs for the voice feature and voice cloning for the familiar parent/guardian.

Challenges we ran into

We definitely ran into lots of problems especially connecting the database to the watch, and making the watch usable.

Accomplishments that we're proud of

A closed loop on real hardware. The watch senses rising agitation, buzzes a breathing cue, and starts calm mode by itself. It then measures whether the calm moment actually helped and feeds that back into its own settings. We tested this end to end on a LilyGO T-Watch S3 Plus. It learns per child. Each child gets their own movement and noise baseline, sensitivity that tunes itself from teacher “false alarm” ratings, and a Thompson-sampling scene picker. Its reward combines the teacher’s rating, the child’s own check-in, and the measured drop in agitation. A full product in one weekend: firmware; an iPhone app (QR pairing, live graphs, insights, location); a Supabase backend with row-level security on every table; a website; AI calm profiles and weekly reports; a calm voice that can be the parent’s own; iMessage texting for parents and teachers. Privacy from the start. Children are known by nickname only. The watch never records audio. Names are stripped before anything reaches the AI. Location is opt-in, and teachers can only see it during school hours. Old data is deleted automatically. We also fixed every issue our security reviews found before demo day.

What we learned

Hardware is humbling. HTTPS failed until the watch synced its clock. Venue Wi-Fi was 5 GHz only, and the watch needs 2.4 GHz. A volume setting silently wrapped around to zero. The tiny speaker sets a hard ceiling on volume. Simple statistics carry a lot of weight. A per-child baseline plus z-scores beat any one-size-fits-all threshold. Defaults matter for kids. Auto-start is off until a parent turns it on, and adults always keep control. Security reviews catch real bugs. Ours found that an email-style iMessage handle could pose as a parent, and that a teacher could switch off auto-start by flipping “false alarm” on and off. Both are fixed. Integrations shape architecture. Photon’s iMessage connection needs a constantly running program, so we built a small bridge with a message outbox instead of forcing it into serverless functions.

What's next for Ebb

Pilot with real classrooms and calibrate the agitation model with special-education teachers, behaviour analysts and occupational therapists. Add heart rate and skin response on future hardware for earlier, more accurate warnings, plus a louder speaker. Report against IEP goals so progress data supports each student’s individualized education plan. Become compliant: FERPA and COPPA review; school data agreements; deleting a child’s data with one tap; encrypted offline storage. Build an Android app and multi-classroom dashboards for schools. Train an on-device model on labelled moments, so the watch spots patterns before the numbers rise.

Built With

• claude
• database
• lovable
• python
• supabase