Scientific Abstract & System Specification
Traditional HSSE incident reporting relies on shallow text tagging and retrospective analysis after accidents occur, missing latent barrier degradation patterns and fatal precursor chains. ASCENSION introduces a 14-layer intelligence architecture combining domain-adapted SafetyBERT models, critical barrier ontology matching via RapidFuzz & pgvector, XGBoost Platt-calibrated dual-threshold classification, and Bayesian Beta-Binomial precursor density shrinkage.
Core Architectural Challenge
Life-saving HSSE precursor intelligence engine connecting hazard precursors across years of industrial history to prevent catastrophic events.
Mathematical Basis & Formal Invariants
Bayesian Beta-Binomial shrinkage, Platt calibrated classification thresholds, multi-criteria decision analysis (MCDA).
Key Empirical Findings & Benchmark Metrics
[01]
Extracted structured barrier states with >93% model confidence across real-world incident narratives.
[02]
Surfaced non-linear precursor recurrence patterns across multi-year field operational logs.
[03]
Live production deployment accessible at ascension.matrikaregmi.com.np.
Verification & Telemetry Testbed
OIL HSSE Industrial Incident & Near-Miss Records
Formal Verification & Engineering Stack
Ayaan SaifiMatrika RegmiNashra ZehraAlisha KhanQayad AliMohammad Aves