Concept brief & pressure-test · v0.1 · August 2026 · for co-founders, investors, and family interviews
Companion artifact: clickable prototype of one live lesson →
A homeschool teacher, usually a parent, sometimes a tutor or co-op instructor, describes a lesson (“the solar system, age 10, 60 minutes”). AI composes it: segments and timing, talking points for the teacher, and interactive content for the student, an explorable orbital model, embedded questions, a manipulable timeline. In the live session, both people share one synced lesson state but see role-specific renders: the teacher gets a conductor view (flow, pacing controls, a live read on each student); the student gets the stage (touch it, spin it, answer inside it). Not a mirrored screen-share, one state, two renders.
Why now: the US homeschool population roughly doubled after 2020 and settled well above its old baseline (~3–4M students); market analysts size homeschooling at roughly $39B in 2026, growing ~11%/yr. Meanwhile LLMs crossed the threshold where they can reliably fill structured content templates, which, as §6 argues, is exactly the safe way to “generate” interactives. The buyer and the teacher are the same person, so the product can sell on teaching experience rather than district procurement.
Each ingredient exists somewhere. The combination, aimed at homeschool, does not:
| Product | What it is | Live human-led sync | AI-built content | Rich interactives | Homeschool-native |
|---|---|---|---|---|---|
| Nearpod / Pear Deck | Teacher-paced synced slides, polls, some sims, district product | ● | ◐ | ◐ | ○ |
| Curipod | AI generates interactive slide decks (polls, word clouds) for classrooms | ● | ● | ○ | ○ |
| Khanmigo | 1:1 AI tutor + teacher planning tools | ○ | ● | ◐ | ◐ |
| Outschool | Marketplace of live classes, taught over Zoom (mirrored screen) | ● | ○ | ○ | ● |
| Time4Learning, K12 | Async homeschool curriculum, student works alone | ○ | ○ | ◐ | ● |
| Synthesis | Cohort game/simulation sessions; AI math tutor, fixed content | ◐ | ○ | ● | ◐ |
| Orbit | AI-composed interactive lessons, conducted live by the parent-teacher | ● | ● | ● | ● |
The defensible wedge is three things working together:
1. Role-specific rendering of one shared state. Everyone else either mirrors a screen (Zoom, Outschool) or paces slides (Nearpod). Rendering the same lesson state differently per role is an architectural choice competitors would have to rebuild around.
2. AI that composes interactives, not slides, safely, by parameterizing a registry of vetted interactive templates rather than generating code (§6).
3. Conductor telemetry built for a parent teaching 1–4 kids, signals become coaching moves (“reopen the Venus card together”), not a 30-seat dashboard. Nearpod assumes a classroom; homeschool tools assume no live teacher at all.
Five vetted interactive templates (explorable model · manipulable timeline · labeled diagram · step-through sim · quick check). AI lesson composer that fills templates and writes talking points + questions, behind a teacher approval gate. Conductor view (flow, pacing, talking points, interaction-only pulse). Student stage. One teacher, up to 4 students. A seeded library of ~20 pre-vetted lessons. Post-lesson recap. Desktop/tablet web only.
Camera/mic signals. Free-form 3D generation. Marketplace. Gradebooks and records (CSV export only). Native mobile apps. Voice chat (families already have a call running or are in the same room, don’t rebuild Zoom).
Validation plan: 10–15 families plus 1–2 co-ops, 6 weeks, 2 lessons/week. The seeded library de-risks AI quality during the pilot; generation is the second act. Metrics that matter: lesson completion rate; teacher edits per AI plan (trust proxy, falling is good); “ran a second lesson” rate; unprompted student requests to do it again; stated willingness to pay at three price points. Two founders can plausibly reach this pilot in 10–12 weeks on the stack in §7.
The riskiest assumption isn’t technical: it’s that parent-teachers want to conduct live lessons rather than hand the kid something self-paced. Everything else is buildable; validate that first, it decides whether the conductor view is the product or a feature.
The research is unambiguous about direction: interaction telemetry beats camera inference for young learners. Expression-based “engagement detection” correlates weakly with actual engagement (studies find facial expressions track only the emotional component, weakly, and even trained humans agree with each other only ~70% of the time when labeling engagement from video). Consumer-webcam gaze tracking adds 100–200px of error, needs calibration a 9-year-old will not sit through, and degrades with glasses, lighting, and wiggling. Meanwhile the tablet already knows exactly what the student touched, when, and how long they hesitated.
| Signal | Reliability for “stuck / disengaged” | Cost & trust risk | Verdict |
|---|---|---|---|
| Wrong-answer patterns (repeat misses, systematic errors) | ● High, directly meaningful | None, already collected | Build first |
| Hesitation latency (time-to-first-action after a prompt) | ● High for confusion; calibrate per kid | None | Build first |
| Navigation shape (wandering vs. rapid-fire vs. idle) | ◐ Medium, ambiguous alone, strong combined | None | Build first |
| Student self-report (“how’s this going?” one-tap) | ● High when framed as help, not test | Trivial; builds agency | Underrated, build |
| Face-in-frame presence (on-device, boolean) | ◐ Medium, catches “walked away” only | Moderate, first camera step | Tier 1, opt-in |
| Webcam gaze tracking | ○ Low on consumer hardware + moving kids | High, biometric-adjacent | Don’t build yet |
| Facial-expression “engagement” scoring | ○ Low, tracks emotion weakly, not learning | Highest, biometric, mislabels kids | Don’t build |
Making signals actionable without making kids feel watched. Three rules. (1) Signals reach the teacher as low-confidence coaching prompts, “Leo may be stuck on Q2: two wrong tries, long pauses. Try reopening the Venus card together”, never as scores, grades, or red badges on the student’s screen. (2) The system may also act quietly: auto-offering a worked example after a second miss reads as kindness, not surveillance. (3) Radical transparency to the student: a “what my teacher sees” panel, one tap away, listing every signal in kid language (the prototype demonstrates this). Kids who can inspect their own data stop wondering about it.
The camera ladder (our lean-in, staged). Tier 0, interactions only; ships in MVP; gets ~80% of the value. Tier 1, on-device face-in-frame boolean, opt-in per family per session, camera indicator always visible, video never leaves the device, only yes/no flags sync. Tier 2, richer on-device signals as research mode, only with per-study consent and a kid-visible switch. Gate each tier on demand from Tier 0 users, not on roadmap ambition: if nobody asks for Tier 1, the interaction data was enough.
COPPA got teeth for exactly this product. The FTC’s amended COPPA Rule (effective June 2025, full compliance required since April 22, 2026) now defines biometric identifiers, face templates, voiceprints, gaze patterns, as children’s personal information; requires separate verifiable parental consent before disclosing kids’ data to third parties (including for ads); and requires a published data-retention policy with actual deletion timelines. Illinois’ BIPA adds a private right of action for biometric capture; Texas and Washington have cousins. Design consequence: raw video and audio never leave the device, ever, any camera-tier inference runs locally and syncs only derived booleans; retention defaults to session-scoped with automatic deletion; no third-party ad tech in the kid surface, full stop.
The homeschool wrinkle cuts both ways. Consent is structurally simple, the parent granting COPPA consent is usually the teacher receiving the data. But two traps: (1) co-op instructors are not parents, the moment a non-parent teaches, you need real verifiable-consent flows, so build them from day one; (2) when parent, teacher, and data-consumer are the same person, the child has no independent advocate, which is why the kid-visible data panel is an ethical requirement, not a feature.
Stored: lesson plans; quiz answers & outcomes; per-segment participation summaries (“explored 6 planets, 2/2 on quiz”).
Transient (discarded after session): raw click/hesitation streams once aggregated; sync event logs after snapshot.
Never collected off-device: video, audio, face/gaze data, keystrokes outside answer fields.
6.1 Keeping two role-views in sync. Split the traffic into three channels with different guarantees, rather than one “realtime” pipe:
| Approach | Strengths | Weaknesses | Call |
|---|---|---|---|
| Authoritative server + ordered event log (WebSocket fan-out) | Teacher authority is natural; reconnect = snapshot + replay; easy to reason about; 50–150ms is plenty for pacing | Server hop on every action; needs regional rooms at scale | ✓ MVP choice |
| CRDT shared document (Yjs-style) | Great offline merge; proven libraries | Peer-symmetric by design, role authority (teacher wins) must be bolted on; lesson state is a state machine, not a document | Not the shape of this problem |
| WebRTC peer-to-peer data | Lowest latency; no relay cost | NAT pain on home networks; no authority; late-join/reconnect is hard | Only for future A/V via an SFU |
Concretely: a tiny control channel (segment changes, pause, spotlights, ordered, authoritative, must-deliver); a lossy batched telemetry channel (student interactions, aggregated client-side before send); and later a separate media plane. Local echo everywhere, the student’s planet spins instantly, the event syncs behind it. A dropped student reconnects by fetching the latest state snapshot plus the event tail, and their interactions buffer locally meanwhile, the lesson never blocks on them. This is also the privacy architecture: telemetry aggregation happens on-device, so the server only ever sees summaries.
6.2 Rich content on weak hardware. A lesson is data, not code: “orbital-model, 8 bodies, these params.” Each template ships three renderers, WebGL, Canvas 2D, and static-with-captions, and a capability probe picks per device. Same schema, same quiz hooks, same sync events; a 2015 laptop gets the 2D solar system and loses nothing pedagogically. (The companion prototype’s stage is the Canvas 2D tier.) This is the answer to “low-end devices homeschool families actually own,” and it falls out of the template registry for free.
6.3 Validating AI content before a child sees it. Five gates: (1) AI fills vetted templates, it never emits runnable code, and physics/rendering correctness lives in the template, not the model. (2) Schema validation rejects out-of-range parameters. (3) Quantitative facts resolve against curated reference data (NASA factsheets, not LLM memory); unverifiable claims get flagged to the teacher. (4) The teacher approves a readable diff before anything goes live, legally and pedagogically, the teacher is the editor. (5) Every element carries provenance plus a one-tap “report an error” that feeds the eval set. Factual QA is a forever-cost: staff it, don’t assume it away.
MVP (pilot in ~10–12 weeks): Next.js + TypeScript on Vercel; Postgres + auth via Supabase; realtime rooms on PartyKit or a single Node WebSocket service (a lesson room is one process, trivially shardable later); templates in three.js/Canvas behind one schema; LLM lesson composition via structured outputs against the template JSON schemas; on-device aggregation in the client. Boring, cheap, two-person-sized.
Scale changes: regional room servers (Cloudflare Durable Objects or Fly) for latency; event log onto a real stream (Redis Streams/Kafka); a content pipeline with human review queues and automated eval harnesses for generated lessons; template SDK for third-party authors; SOC 2 + COPPA safe-harbor certification; native tablet clients.
Adoption (highest): the conductor assumption, parents may prefer self-paced content over conducting live (§3). Homeschool is also several markets wearing one label (faith-based, secular-academic, unschooling), each buying differently; and price sensitivity is real. Mitigation: pilot across segments; co-ops as beachhead, one instructor decision reaches many families and justifies subscription pricing.
Ethical/trust: a false “disengaged” flag lands harder at home than at school, the teacher is the parent, and dinner follows the lesson. Mitigation: confidence-framed nudges, no scores, kid-visible data, camera tiers off by default. Trust is also the moat: the first bad press story about “AI watching homeschooled kids” poisons the category.
Technical: AI lesson quality variance (mitigated by templates + seeded library); sync robustness on rural bandwidth (event-sourced design degrades gracefully); template breadth, five great templates beat fifty mediocre ones.
Market: Nearpod bolting on AI, or Khanmigo adding live multiplayer, reaches adjacent ground. Neither is homeschool-native nor conductor-shaped; speed in the niche plus the template library is the defense.
1. Wedge user: solo parent-teacher, hired tutor, or co-op instructor first? (Changes onboarding, pricing, and consent flows.)
2. Is live-first right, or does the same lesson need a self-paced replay mode as the daily driver, with live as the weekly anchor?
3. Content strategy: how big must the seeded, human-vetted library be before generation becomes the default path?
4. Subject order after science? (History timelines are easy; math sims are the hardest to template well.)
5. Pricing: family subscription, per-lesson credits, or co-op license, and what does the pilot say about willingness to pay?
6. What does the recap loop look like when the teacher is the parent, who is the report for?
7. Which nudges do teachers act on in practice? Instrument the pilot before building more inference.
8. Does anyone ask for the camera tier unprompted? Gate that investment on observed demand.
9. Account model under COPPA: parent-managed child profiles only, or child logins with verifiable consent?
10. What is the lesson-quality rubric and golden set for evaluating generated content before scaling it?
FTC, Children’s Online Privacy Protection Rule amendments (2025): ftc.gov/legal-library/browse/rules/childrens-online-privacy-protection-rule-coppa · Homeschooling market sizing: Custom Market Insights, “Homeschooling Market 2026–2035” · Expression-based engagement detection reliability: peer-reviewed studies incl. Frontiers in Psychology (2023) on facial expressions vs. self-reported engagement · Product references: nearpod.com, curipod.com, khanmigo.ai, outschool.com, synthesis.com, time4learning.com · Planetary data: NASA planetary factsheets (nssdc.gsfc.nasa.gov).