Deep Blue designed a layered AI architecture combining agentic orchestration, generative explainability, and deterministic optimization – delivering AI-assisted maintenance planning that meets aviation-grade governance.
AI for the most constraint-intensive scheduling problem in aviation
Challenge
Aircraft maintenance planning requires continuous coordination of utilization schedules, regulatory compliance windows, labor certifications, station capacity, parts and tooling availability, and real-time disruption events including AOG scenarios.
Trax needed AI that could meaningfully improve planner productivity and trust while preserving deterministic decision authority, full auditability, and regulatory defensibility. Autonomous decision-making was explicitly out of scope by design.
- AI cannot modify schedules or compliance outcomes directly
- Human planner approval required before any plan commits
- Multi-tenant isolation – zero cross-tenant data exposure
- Every AI explanation traceable to deterministic scoring
- AOG replanning in ≤1 minute with zero constraint violations
Solution: Four-Layer Governed Architecture
Agentic orchestration drives workflow execution. Generative AI provides explainability. Deterministic engines retain full authority over scheduling and compliance outcomes.
MAIN PLANE
- Trax eMRO integration services and existing workflows.
DETERMINISTIC PLANE
- Constraint optimization, rules engines, compliance logic – the system of record.
AGENTIC AI
- Orchestration of planning, replanning, assignment & utilization workflows.
GENERATIVE AI
- Bedrock explanations & NL interface – constrained to deterministic outputs.
Results: Responsible AI & Operational Discipline
Governance & Responsible AI
Deterministic logic remains system of record for all scheduling outcomes. No autonomous execution -planner approval required for every plan. Zero cross-tenant data sharing. Prompt templates version -controlled in Git. Field-level data classification governs AI data exposure.
Operational Discipline
100% of agent actions and Bedrock invocations logged via CloudWatch. Human approval gate enforced on all planning executions. Override capture rate ≥95% with full reason and context. IAM least-privilege enforced for all AI service identities.
Agents orchestrate. Bedrock explains. Deterministic logic decides.
Agentic Capabilites
- Planning orchestration agent
- Event-driven replanning agent
- Technician matching agent
- Utilization tracking agent
- Human-in-the-loop control
MVP Features Delivered
- Constraint-based optimizer
- Smart work pack generator
- AOG disruption replan bot
- Technician matching engine
- NR analysis forecasting
