Mission Brief
Thermal Anomaly Root Cause Analysis
Design an AI system to identify root causes of temperature oscillations in nuclear reactor cooling loops
Primary Objective
Design and implement an AI system that can perform rapid pattern matching to identify historical instances of similar temperature oscillation patterns across 40 years of data, build multi-modal root cause analysis with agentic reasoning, implement physics-informed validation, and generate regulatory-grade traceability documentation.
Secondary Objectives
Predict if the anomaly will self-correct or worsen
Recommend immediate operator actions with safety justification
Estimate time window before mandatory NRC notification
Identify preventive measures for future operations
Technical Requirements
Document Processing & OCR
Extract data from degraded 1980s-era scanned documents
Handle handwritten control room logs with multiple operators' writing styles
Parse engineering drawings and P&IDs to extract component relationships
Reconstruct time-series data from tabular scan images
Agentic Search & Reasoning
Autonomous query generation to explore related systems
Deep-research through maintenance histories to find similar component issues
Cross-document correlation (e.g., 'valve V-203 maintenance' in work orders → 'Loop B temperature spike' in logs)
Hypothesis generation and evidence gathering loops
Multi-Modal Integration
Combine time-series sensor data with text-based maintenance logs
Analyze thermal camera footage alongside written observations
Correlate acoustic signatures with historical pump performance data
Quality Control & Verification
Human-in-the-loop checkpoints for critical findings
Confidence scoring for each causal hypothesis
Physics-based sanity checks on proposed mechanisms
Anomaly detection on AI outputs (hallucination prevention)
Success Criteria
Technical Performance
Accuracy: Identify root cause confirmed by subsequent physical inspection
Speed: Complete analysis in < 4 hours (50% faster than traditional engineering team)
Recall: Find all 3+ historical precedents buried in 40 years of records
Traceability: 100% of conclusions traceable to source documents with page numbers
Regulatory Compliance
Documentation meets 10 CFR 50 Appendix B quality standards
Reasoning chain suitable for NRC review
Uncertainty quantification on all recommendations
Operational Impact
Operators can understand and act on recommendations
No false positives that trigger unnecessary plant shutdown
Recommendations prevent recurrence
Evaluation Dimensions
Architecture Design
30%How do you structure the agentic pipeline? What models/techniques for each component? How do you handle the 'cold start' problem with unfamiliar data?
Data Engineering
25%OCR strategy for low-quality scans, time-series alignment across different eras, schema inference for unstructured logs
AI/ML Approach
25%Pattern matching across heterogeneous data types, causal reasoning under uncertainty, physics-informed constraints
Verification & Safety
20%How do you prevent hallucinations on critical safety decisions? What human checkpoints are essential? How do you quantify confidence?
Discussion Prompts
Tradeoffs: Speed vs. accuracy when lives are on the line
Edge Cases: What if no historical precedent exists?
Scaling: How would this system work across a fleet of 50 reactors?
Failure Modes: What could go wrong with your approach?
Regulatory Strategy: How would you convince NRC this AI-assisted analysis is trustworthy?
Deliverables
15-Minute Presentation
System architecture diagram
Key technical decisions and rationale
Mock output showing traceability chain
Risk mitigation strategy
Code sample (optional): Pseudocode or actual implementation of one critical component
Bonus Challenges
Design the monitoring system to catch this BEFORE it becomes an anomaly
Propose an active learning approach to improve with each event
Sketch the UI for operators interacting with AI recommendations during crisis
This mission brief tests your ability to build AI systems that nuclear operators will trust with safety-critical decisions during time-sensitive situations—the heart of what Gordian does every day.