A HIPAA-aware clinical mental-health education platform, in active beta deployment - curriculum researched against current clinical literature using deep research and clinically validated by board-certified psychiatric nurse practitioners at Real Psychiatric Services, protected by the MAIA distress-classification safety architecture.
Behavioral-health licensees - psychiatric practices, multi-clinic franchises, behavioral-health benefits providers, university counseling centers - operate DWA as their credentialed digital-education arm.
Every AI ingress is gated by MAIA - sentinet/suicidality, ELECTRA-base, C-SSRS clinical training. On the first euphemism we do not fire an alert. The coach enters screening mode and asks one open-ended calibration question. On heavier disclosure the alert fires - one row, one episode, even if MAIA fires three times across screening turns. The provider sees a pre-computed payload: anonymized alias, transcript with MAIA classifications inline, urgency tier, lesson suggestions frozen at fire-time, PMHNP-curated response templates.
Architecture: screening state machine (deriveState + decide as pure functions) in lib/safety/episode.ts · alert dedup keyed on (patient, session) in lib/safety/session-store.ts · lesson suggestions from retrieveLessonsForCoach in lib/ai/rag.ts · 7 PMHNP-curated response templates seeded at boot · production-grade data model and state machine; alert detail composer is demo-grade UI on production endpoints; patient-side care_message rendering ships in open beta. Full architecture in docs/positioning/screening-and-alert-architecture.md.
A weaker gate would have promoted v1 on its in-distribution F1 alone and shipped a worse-than-baseline experience. Our gate caught it. Nebius compute is what feeds the gate the production-class candidates that earn it - LLaMA-3-8B coach LLM, multilingual MAIA, per-tenant fine-tunes.
services/maia/validation/.The promotion gate runs on Apple Silicon today. Nebius compute scales the gate's candidates - the canonical Layer-6 in-house coach LLM, multilingual MAIA retraining cadence, per-tenant fine-tune capacity.
Per 20-turn coaching session. Same output, two architectures. Reproducible from unit-economics-model.csv using vendor-published pricing - re-runnable in any audit. The DWA wellness-session counterpart runs at 4.4x today; the Layer-6 ambitious projection assumes the in-house coach LLM Nebius unlocks absorbs 60% of heavy reasoning.
SVTech Consulting Services LLC requests Nebius compute credits to deliver Layer 6 in production, scale MAIA retraining cadence, and unlock multilingual coverage for the documented English-only limitation. DWA in active beta at digitalwellness.academy. Open beta May 2026 acquires the marquee at-scale Stage-1 beta partner. Worldwide-when-localized: NHS, EU AI Act, APAC equivalents are manifest variants, not codebase forks.