An EdTech company ships an AI-avatar K to 12 LMS in 12 weeks

Challenge

• Single-teacher classrooms cannot scale instruction. Demand for personalised, culturally relevant content outstrips what any one teacher can produce.
• K to 12 platforms have to be FERPA and COPPA compliant from day one. Bolt-on compliance loses funding and trust.
• Avatar teachers usually require GPU servers per concurrent student. The cost model breaks the business.
• Students have to feel met by their teacher. Generic AI loses the room in five minutes.

Approach

• Browser-based 3D avatars on an MIT-licensed Three.js / WebGL pipeline. Lessons render on the student's Chromebook with zero GPU server cost.
• Multi-LLM gateway in front of OpenAI, Anthropic, and Google. PII is stripped before any external call. Per-task routing keeps cost predictable; per-student spend caps.
• Adaptive conversational loop. The avatar slows down, re-phrases, or asks a different way when a student stumbles.
• Compliance baked into the architecture. FERPA, COPPA, BIPA, SOPIPA, and NY Ed Law 2-d treated as design constraints, not a pre-launch checklist.
• Avatar art sourced via an open 3D pipeline for full ownership and culturally diverse representation. No license traps.

What shipped

• AWS foundation in the client's own AWS account, VPC, secrets, object storage, managed database, baseline encryption and RBAC at the API gateway.
• Three role-scoped portals (Student, Teacher, Admin) with global search and filter.
• Course and lesson management with per-lesson avatar assignment, script upload, and preview workflow.
• AI Avatar Engine with lip-sync, 52 ARKit blend shapes, eight moods, and gesture support, running entirely in the browser.
• TTS in-browser, zero cost for English. STT via Whisper.
• Conversational AI with adaptive pacing, multi-LLM routing, and misunderstanding detection feeding back into avatar pacing.
• Multi-step student enrollment with guardian relationship and lifecycle status.
• Teacher KPI dashboard, struggling-student alerts, per-avatar engagement, time-per-grade, course stats.
• Compliance foundation, encryption, RBAC, audit trail, content safety filter on AI output, kill switch on avatar output.

Outcomes

• MVP live with at least one pilot classroom by end of Month 3.
• Avatar-led lessons running on school Chromebooks without a GPU server.
• Teacher authors, previews, and publishes a lesson with an assigned avatar in one sitting.
• Student enrolled with a guardian, attending an avatar lesson, with KPI dashboards reflecting live classroom data.
• All student data resident inside the client's AWS VPC. No PII leaves the perimeter without an approved DPA path.

Stack

• AWS (VPC inside the client's account)
• Next.js
• browser-based 3D avatars (Three.js and WebGL)
• in-browser TTS
• Whisper
• multi-LLM gateway over OpenAI
• Anthropic
• and Google
• PostgreSQL
• S3
• IaC for dev and prod

What's next

A planned follow-on (Sprints 7 and 8) adds back the features trimmed for the Lean MVP. Phase 2 layers gamification, parent alerts, IEP and Section 504 accommodations, and district branding. Phase 3 introduces predictive analytics, early warning, and multi-district deployment.