Rebound

Plans fall apart. Rebound helps you bounce back. Built for ADHD brains, it turns overwhelming schedules into achievable steps and lost time into learning opportunities. Enabled by Bee data + AWS.

Comment

"If it's not in front of me, it doesn't exist."

We both have ADHD, and we've both worn Bee for weeks: lectures, study sessions, whole days. Bee became our external memory. It knew what the professor said, what we promised a teammate and when the problem set was due.

But planning was never our problem. Recovering when the plan breaks is. A deadline said out loud in lecture is gone by the time we reach the door. A meeting runs over by 20 minutes and the afternoon collapses, because re-planning takes the exact executive function that just ran out. And a 90-minute lecture fades within days unless we come back to it, which is the hardest thing of all to start.

Every ADHD planner waits for you to notice the plan broke. That noticing step is exactly the step ADHD takes away. Bee already hears the moment the day goes sideways, so we built Rebound to act on it.

What it does

Rebound is an iPhone app that extends Bee for students with ADHD: a personal tutor that lives in your pocket. Bee hears your day. When it breaks, Rebound re-plans it. Then it turns the time that frees up into learning that fits you: quizzes, a 90-second explainer video and a podcast episode, all made from your own classes. You keep using Bee; Rebound works on top of it.

🔁 1. Re-plan: the day breaks, Rebound bends

"Oh, my class got cancelled. I don't know what to do with my free time." Bee hears it. Rebound frees the hour, slides the right block in (the problem set due tonight), gives exactly one next step, and writes the new reminder back to your Bee todo alarms, with one-tap undo. It also catches meetings that run over, late starts, drifting and a spoken "where was I?". You never open an app to re-plan.

🧠 2. Learn: a quiz from the lecture you just heard

On Friday a data scientist from Meta gave a guest lecture in our ML class. Bee heard it and kept the slides we photographed. Rebound turned it into a quiz in the app:

  • Six formats, including tap the answer on your own slide photo.
  • You say if you're sure first. A confident surprise is the one that sticks, so it comes back tomorrow as "ready to refresh" (a Bayesian estimate per key idea, an FSRS schedule per concept).
  • A checker in code drops weak questions, so a round is never empty or off-topic.

🎬 3. Recaps: a review video for what didn't stick yet

For an ADHD brain, a 90-minute lecture is a wall. A 90-second explainer is a door: short, visual, captioned and made from your own class, so starting is easy. Recap Studio, an agent on Amazon Bedrock AgentCore, makes it from the same Bee data the quizzes use:

  • Claude on Bedrock writes the script from fixed Jinja prompts, with your slide photos as reference.
  • Nova Sonic on Bedrock narrates. Remotion assembles the scenes and ffmpeg does the final cut inside the container.
  • The video pauses for a 5-second quick check, so it's review, not just watching.
  • Every stage is checkpointed, so a recap always finishes. A separate token gate, a daily cap and its own dashboard keep it at cents per recap.
  • The same narration ships as a podcast episode.

📺 4. Fire TV: watch it at home

Rebound TV, built on Amazon's Vega video sample, shows your recaps in a Your recaps row and plays them with the quick-check pauses. Every chapter has its own deep link (rebound:// on the iPhone, rebound-tv:// on Fire TV), so a reminder from your Bee opens the exact moment you need, on your phone or your TV.

🔋 Paced to you

Every student learns differently. Bee's facts show how you like to learn, Apple Health shows how you slept (sleep, HRV), and Rebound's agent keeps an eye on both. Today that might be three practice problems, or one 90-second video.

🤖 Also: Alexa, MCP and open agent skills

  • Alexa: a Rebound review skill plays your latest recap as a podcast episode (built and tested; not yet certified).
  • Rebound MCP server with 14 tools, next to Bee's own MCP server, for Claude, Codex, Cursor or your own agent. Writes need a second "yes" and can be undone.
  • Open agent skills bee-quiz and bee-recap: the same engine with no AWS, for anyone with a Bee and Claude Code or Codex.

💛 Designed for ADHD

One Now item. At most three open loops in sight. A Circuit Breaker that shrinks everything to one kind step. Copy that never says "missed", "late", "overdue" or "wrong". No streaks, no red.

How we built it

  • Hear: bee stream on ECS Fargate → API Gateway → an ingest Lambda that catches cues like "my class got cancelled" → DynamoDB.
  • Understand & teach: a graph Lambda pulls transcripts, voice notes, Bee photos and facts from the Bee API; Claude Haiku 4.5 on Bedrock turns them into concepts and practice items and reads slide photos with vision; a quest builder sizes rounds by how you slept.
  • Re-plan: Claude Sonnet 4.5 on Bedrock re-plans the hour, and the Bee API syncs the reminder back to Bee.
  • Recap: Recap Studio on Bedrock AgentCore Runtime (an arm64 container in ECR): storyboard → voice → visuals → render → publish, checkpointed in S3, served through CloudFront signed links to the iPhone, Fire TV and the podcast feed. Ready recaps create Bee to-dos that carry the chapter deep link.
  • Surfaces: a native SwiftUI iPhone app (HealthKit, Live Activity, widgets), Rebound TV on Vega OS (React Native for Vega), an Alexa skill, an MCP server and two agent skills.
  • Production-grade: all infrastructure in Terraform, CloudWatch dashboards and alarms, a $50 budget with an automatic killswitch, and 1,700+ automated tests (1,055 backend, 44 Recap Studio, 391 Rebound TV, 200+ iOS).

It runs on real data. The live cloud pipeline turned Friday's guest lecture into a 94.7-second recap in about 8 minutes for about 11¢: Claude wrote it, Nova Sonic voiced it, and it plays on the iPhone and in the Rebound TV catalog.

Privacy by architecture

  • No audio, ever. Transcript text is used in memory and discarded.
  • No photo bytes stored. Photos are read once; photos of people are dropped.
  • Health stays on the phone. Only daily numbers are posted.
  • The Bee token lives only in Secrets Manager and the iOS Keychain. Recap media is private in S3 and reached only through short-lived signed links.

Challenges we ran into

  • Bee's private root CA broke every server-side client until we found and pinned it.
  • An at-most-once stream with no event ids, so we built our own gap-filling and idempotency.
  • Bee's voice assistant wasn't reliable enough to drive Rebound, so Rebound listens to the ambient transcript and catches cues itself.
  • Alexa+ can read Bee but can't update it. We asked Alexa to correct where a meeting happened. It said it had "forwarded that correction to Bee", twice, then confirmed "the correction hasn't been applied yet."

All of it is in our Bee developer friction log (23 items), each with the fix we'd suggest.

What we learned

  • For ADHD, the recovery is the product. One step beats a list.
  • A wearable's value is the context it hears. The craft is turning it into one action at the right moment, then getting out of the way.
  • Let the model write and let code judge. That made the same engine portable from Lambda to AgentCore to a laptop skill.
  • Running agents in production means killswitches, token gates, checkpoints and idempotent pipelines.

What's next

  • Recap variations instead of full regeneration: re-voice or re-cut one scene, a shorter or slower take, or a different analogy, reusing the checkpoints the pipeline already writes.
  • Graph-based indexing of what's on the slide: figures, equations, diagrams and code become interactive learning widgets, and they feed a personalized, Duolingo-style curriculum.
  • A shared knowledge graph across educators and learners, so a concept learned once is reused in future courses.
  • A secure, autonomous AI tutor that keeps improving itself and gets more efficient with every learner.
  • Recursive memory. In a prototype we tested, it cut repeated multimodal inference by up to 87.5% (64 → 8 vision-model calls) while keeping 100% recall on previously learned concepts. We didn't ship it yet because of time.
  • Per-student accounts (Cognito) and a per-user Bee token vault, so Rebound scales from one student to a whole campus.
  • Certify the Alexa skill, ship Rebound TV to the Appstore for Vega OS devices, and publish on the App Store.

Roughly 6.1 million kids in the US have ADHD, and everyone benefits from "what did I say I'd do?"

Bee hears your day. Rebound helps you rebound, and excel, in life.

Built With

Share this project:

Updates

Submission history