Data and AI processing
What Praxis sends to models, and what it does not do with that information.
Praxis can draft evaluation sections, goal language, and service recommendations from evidence the therapist has entered or confirmed. Generation is bounded by the current evaluation context. Unconfirmed or rejected prior-report extractions are excluded.
When a live model is connected
Selected evaluation context is sent only to the live provider or providers explicitly enabled by the operator. Multi-provider fallback is disabled by default so a timeout does not automatically forward the same student context to every configured vendor. Supported integrations include Groq, Cloudflare Workers AI, OpenAI, Anthropic, OpenRouter, and Google Gemini. OpenAI requests disable provider-side storage, and OpenRouter requests require zero-data-retention by default. Provider keys stay server-only.
When no live key is configured
The labeled sample workspace uses a deterministic mock provider and does not present mock output as a live-model result. Authenticated live accounts do not fall back to mock generation.
Guardrails
Drafts are checked so entered scores are not rewritten, missing scores are not invented, missing baselines are not fabricated, suggested targets stay labeled, and service rationale does not silently invent frequency or duration. Diagnosis, eligibility determinations, and claims that occupational therapy is legally required are blocked.
Human review
Therapists review, edit, approve, and finalize. Exporting a PDF records an audited export.Praxis does not submit documents to a district IEP system.