CulturalBERT-VLAP is a clinically-validated NLP model built on culturally-specific language data from BIPOC, LGBTQIA+, and underserved youth communities. It catches what generic AI misses.
Standard sentiment models are trained on majority-White internet text. They systematically misread AAVE, code-switching, queer vernacular, and trauma-adjacent language. VLAP was built to close that gap.
Expressions like “I can’t keep doing this no more fr” are flagged as hopelessness markers — not penalized as informal or dismissed as non-standard.
“It’s not that deep but” is a documented pre-disclosure pattern in youth language. VLAP detects it as signal CCM-04 and increases contextual weight on what follows.
“Unaliving” is a youth community code term for self-harm ideation. It is in the VLAP vocabulary. Standard models have never seen it. VLAP catches it every time.
Sudden positive shifts after distress (“everything is fine now, I’m over it”) are recognized as high-risk farewell patterns (HOP-07), not genuine improvements.
Every piece of member text flows through a 7-step pipeline that scrubs PII, checks consent, runs inference, and delivers structured output — all without storing a single word of raw text.
Text arrives from peer posts, journal entries, or chat
Consent verified. No consent → rejected, never processed
Microsoft Presidio removes all identifying info before inference
CulturalBERT-VLAP classifies risk tier and detects active signals
Structured result delivered to therapist dashboard only. Raw text discarded.
Each signal is a clinically-defined indicator that warrants human review. None are diagnostic — they surface for a licensed clinician to evaluate.
8 signals including direct statements, AAVE constructions, and performative positivity masking.
9 signals including social disconnection, family rejection language, and digital-native unseen expressions.
7 signals including coded language (“unaliving”), indirect method references, and behavioral withdrawal.
6 signals including time-bounding (“just gotta make it through tonight”), farewell patterns, and sudden calm.
12 modifiers including AAVE register, minimization, code-switching, and anti-LGBTQ+ political stressors.
VLAP was built with six non-negotiable principles that govern every design and deployment decision.
No member ever sees a risk score, signal code, or AI flag. Outputs flow exclusively to licensed therapists and authorized administrators.
No automated action is triggered by AI output alone. Every high-risk flag requires clinician review within 24 hours. Crisis flags within 90 seconds.
Member text is processed in-memory and immediately discarded. Only structured inference output is stored. Retention: 90 days. Audit logs: 6 years.
False positive and false negative rates are tracked by demographic subgroup in every inference batch. Systematic disparities trigger mandatory model review.
Our 37-page VLAP V1 Technical Spec covers architecture, API contract, bias controls, and acceptance criteria.