Technology
Multimodal AI engineered for classrooms
Transparent by design: educators see how inputs route, how memory persists, and where human oversight plugs in.
Input processing
- Text: NLU, intent routing, safety classifiers, and structured tool calls.
- Image: Bedrock Claude Vision, optional Amazon Textract for scans, custom rubrics for grading assists.
- Draw: Client-side canvas to PNG; treated as a first-class bitmap for vision—ideal for scratch work without a camera.
- Voice: Sonic speech-to-text, streaming partials, accent-aware adaptation.
- Video: Motion-aware sampling, key-frame extraction, synchronized transcript alignment.
Orchestration & models
AgentCore coordinates tools, memory, and policies. Claude Opus handles deep reasoning, Sonnet balances latency and quality for vision-heavy tasks, Haiku covers rapid retrieval-style answers.
Memory & retrieval
Embeddings power semantic recall across modalities; configurable retention windows meet district policy. AgentCore Memory keeps session state synchronized whether the learner is typing or speaking.
Outputs
Multi-format responses combine formatted text, lightweight overlays on student uploads, Polly-class speech, and optional downloadable practice sets.
Data privacy & security
- TLS everywhere; optional customer-managed keys for storage.
- FERPA/COPPA-aware workflows—finalize policies with your counsel.
- SOC 2 Type II on the roadmap—display badges only when audits complete.
- Media handling: ephemeral processing by default, encrypted at rest when retention enabled.
- No biometric voiceprint storage; transcripts optional per org.
Performance & scale
Global AWS regions, autoscaling inference, and edge caching for static assets. Target SLOs mirror classroom realities—see homepage deep dive for modality latency budgets.
Accessibility engineering
Components ship with logical focus order, ARIA for live regions, captions for synthesized speech, and high-contrast tokens. Pair with your institution’s AT evaluations before production rollout.