Technical transparency for our AI infrastructure.
Last Updated: June 1, 2026
Processes free-text inputs from learners during simulations and generates context-aware, culturally appropriate responses from NPC counterparts in real-time.
Handling learner negotiations, conflict resolution, active listening, and rapport-building inside bounded simulation environments.
Evaluating psychological health, predicting actual real-world promotion viability, or processing highly sensitive PII.
Response Latency < 600ms, Cultural Nuance Accuracy 94%
Arabic/English Parity: 98.5%
Modern Standard Arabic, GCC Dialects, US/UK English.
Responses are logged and audited. Safety filters prevent generating harmful or prohibited content.
"May occasionally default to overly formal Arabic if prompt context lacks specific regional dialect instructions."
Acts as a supportive 'mirror' for learners, analyzing session data to offer constructive feedback, highlight blind spots, and suggest alternative approaches.
Post-session feedback, mid-scenario prompts, explaining trade-offs, and linking behavior to established leadership frameworks.
Making binding employment decisions, replacing human managerial judgment, or disciplining employees.
Helpfulness Rating 4.8/5, Hallucination Rate < 0.1%
Arabic/English Parity: 99.2%
Modern Standard Arabic, English.
Learners can flag unhelpful coaching. All severe flags trigger human review of the session logs.
"Feedback is constrained to the context of the simulation; Leena does not 'know' the learner's actual workplace history."
Analyzes learner interactions to calculate objective scores (Decision Quality, Time-to-Decision, Rapport Index) for HR/L&D dashboards.
Cohort-level reporting, tracking skill progression over time, identifying systemic organizational blind spots.
Automated firing/hiring decisions without human review. The model provides insights, not mandates.
Scoring Consistency 96%, Replay Validation 100%
Metric Generation is Language-Agnostic
Operates on raw interaction data (language-independent).
All high-stakes scores are appealable via the Human Appeal Process. The model logic is periodically audited for bias.
"Scores rely heavily on the defined optimal paths of the scenario; deviations that are creative but undocumented may be initially scored lower."
Due to the proprietary nature of our specific configurations, prompts, and guardrails, full technical Model Cards (including data sets and bias testing logs) are available to enterprise IT and security teams under NDA.
Contact Data Privacy Officer
Inquiries or complaints regarding data processing may be directed to SDAIA or your local authority.
Aligned with Saudi PDPL, designed with NCA ECC controls in mind, and respects international frameworks (GDPR/CPRA).
