Wedge

Target mid-sized chemical refineries plagued by frequent nuisance alarms and subsequent unit trips. This niche provides acute financial motivation and standardized historian architectures for rapid proof of value. Expand from post-incident root cause analysis into real-time operator advisory, eventually capturing predictive maintenance budgets across broader heavy industry verticals.

Timing

Foundational models now possess the context windows and reasoning speeds required to parse massive volumes of industrial event logs and time-series metadata in seconds. Additionally, a generational retirement wave leaves control rooms staffed by junior operators who lack the decades of tribal knowledge previously relied upon to diagnose complex alarm cascades.

Why This ICP

Continuous process facilities like chemical plants and refineries incur massive financial penalties from unplanned downtime and safety incidents. They already centralize their operational data in plant historians, providing immediate and high-fidelity data feeds for integration.

Size Of Prize

Approximately 15,000 continuous process manufacturing and power generation plants operate globally. At an estimated annual spend of $50,000 per plant for advanced alarm management and operator advisory tools, the addressable market totals $750M.

Gap Narrative

Control room operators face overwhelming alarm floods during process upsets, often receiving hundreds of alerts per minute that obscure the root cause of the failure. Traditional alarm management software relies on static suppression rules and fails to dynamically identify the originating fault. Operators require a system that instantly collapses cascading operational alarms into a single actionable incident context.

Defensibility

Defensibility scales through deep workflow lock-in and the accumulation of plant-specific causal graphs. As the system processes historical and live alarm floods, it maps the unique physical interactions of the facility's equipment. This localized intelligence creates extremely high switching costs, as a new generic model cannot replicate the learned equipment relationships without access to the same proprietary operational history.

Why This Thesis

An agentic approach directly mitigates the acute cognitive overload operators experience during emergencies. Rather than presenting another software dashboard to interpret, an active agent parses the flood, suppresses downstream symptom alerts, and surfaces the specific malfunctioning asset requiring immediate intervention.

Hardest Part

The single hardest part is deterministic causality extraction from noisy, unstructured log streams in real-time during an active flood, ensuring the root cause is never suppressed alongside symptomatic noise.

Min Viable Scope

Build ingestion for a single primary alerting tool and output a consolidated incident timeline with top-three probable root causes. Deliberately leave out automated remediation, complex multi-system correlation, and custom suppression rule builders.

Cold Start Problem

Models lack domain-specific topology and historical incident context for a new environment. Break this by ingesting historical ticketing and alert data during a read-only shadow period to train baseline groupings before enabling active suppression.

Time To First Value

2-4 weeks of passive data ingestion

Data Moat Available

true

Technical Difficulty

High

SAM

~$100-150M US and European Tier-1 petrochemical facilities with modern distributed control systems

SOM

~$10-25M

TAM

~4000 global petrochemical and refining facilities x ~$75k/yr = ~$300M

Growth Rate

~8-12%/yr, driven by retiring senior operators, increasing sensor density, and strict process safety management compliance mandates

Paid Comparable Spend

~$50k-150k/yr per facility on periodic alarm rationalization consulting, static alarm management software licenses, and post-incident investigation labor

Kill Thresholds

• DCS integration requires > 14 days of custom engineering
• Operator acceptance rate of triage recommendations < 40%
• Pilot conversion rate < 25% after 60 days
• Sales cycle exceeds 120 days for Tier-1 facilities

Leading Metrics

• Time-to-first-value via read-only DCS integration
• Percentage of secondary alarms automatically clustered during an upset
• Operator acceptance rate of triage recommendations
• Weekly active process engineers reviewing historical flood logs

What Proves Right

Shift operators rely on the product to suppress secondary alarms within 60 seconds of a plant upset. Facilities sign $50k annual contracts from safety management budgets after a 30-day read-only pilot. Process engineers log in weekly to review the clustered root-cause alarm reports.

What Proves Wrong

Plant managers refuse to connect the system to the live Distributed Control System due to read-only compliance rules. Operators ignore the triage interface and continue acknowledging alarms manually in the legacy console. The pilot phase stalls indefinitely waiting for safety committee approval.