SentinelZero: An AI-Powered Multi-Agent Industrial Safety Intelligence Layer for Compound Risk Detection and Pre-emptive Hazard Intervention in Heavy Industry

SentinelZero demonstrates that the compound risk detection problem in industrial safety—where KikTro Labs SentinelZero 2026 Technical Paper | Page 26 dangerous conditions emerge from simultaneous multi-domain signals individually within normal bounds—is tractable using a deterministic, explainable multi-agent architecture without requiring large historical datasets or black-box deep learning models. The system's core innovation is its Compound Risk Fusion Agent and its seven compound multiplier rules, which model documented hazardous signal interactions from industrial accident databases. Applied to a Vizag-inspired scripted incident scenario, SentinelZero achieves compound risk detection 47 minutes before any conventional threshold alarm would activate, with zero false negatives across three scenario types. The architectural decision to use LangGraph for agent orchestration provides a clean, typed, auditable agent workflow that directly reflects the sequential logic of industrial safety assessment. The choice of deterministic risk scoring over LLM-based safety decisions reflects the fundamental engineering principle that safety- critical outputs must be reproducible and independently verifiable. LLM capabilities are applied only to human-facing functions (report generation, compliance Q&A) where creative reasoning adds value without compromising safety determinism. The React/SVG geospatial heatmap, designed around cognitive immediacy principles, enables plant operators to assess compound risk across all zones in under three seconds—a human-factors outcome that is as important as the underlying detection capability. An undetected risk that is not communicated actionably to an operator in time provides no safety benefit. We conclude that SentinelZero's architecture represents a viable path to next-generation industrial safety intelligence: one that is explainable to safety engineers, auditable by regulators, actionable by plant operators, and extensible to production SCADA integration. The prototype demonstrates proof-of-concept for all five hackathon judging criteria—compound risk detection, predictive lead time, geospatial situational awareness, regulatory compliance intelligence, and low false-negative rate—within the engineering constraints of a four-day hackathon sprint. The sentence that should be remembered from any demonstration of SentinelZero: "Every individual system still considered operations normal. SentinelZero identified the dangerous combination and predicted escalation before threshold alarms would have activated."