Deep Blue designed a three-plane architecture that operationalized AI-assisted deal analysis to the same production standards governing every other service on Airvoyant’s TSN platform – zero deployment exceptions, zero out-of-band changes.
The risk wasn’t model quality – it was operational governance
Challenge
Airvoyant’s TSN platform operated with mature DevOps practices: structured CI/CD pipelines managed through Bitbucket, Terraform-based infrastructure as code, and a well-defined environment promotion path from development through staging/UAT to production.
When evaluating AI adoption, Airvoyant recognized that introducing AI as an exception path – outside of pipeline controls, with direct data access, or with ad-hoc configuration practices – would undermine years of engineering discipline and create compliance exposure inappropriate for a regulated aviation SaaS platform.
Multi-tenant isolation, deterministic decision-making, and procurement auditability requirements all constrained how AI could be deployed and operated.
Solution: Three-Plane Architecture
The solution established a clean separation of concerns across three service planes – ensuring AI was treated as a first-class production workload, not an exception.
MAIN PLANE
Existing Airvoyant TSN services and workflows – untouched by the AI integration.
DETERMINISTIC PLANE
Rules-based scoring, hard disqualifiers, weighted evaluation logic, and vendor ranking – the authoritative system of record for all deal analysis outcomes.
AI PLANE
AWS Bedrock Agent Core generates natural-language explanations and recommendations, constrained exclusively to the outputs of the Deterministic Plane.
Structured Delivery Model
A four-month program organized into three phases, each with explicit scope boundaries and acceptance criteria.
- Phase 1: Architecture design, attribute finalization, environment setup, draft validation test cases
- Phase 2: Bedrock Agent Core configuration, MCP server implementation, deal logic development, integration testing
- Phase 3: UX integration support, scenario testing, tuning, stabilization, formal readiness review. Each deliverable carried a defined 5-day accept/reject window.
What we built
- Three-plane architecture design
- Bedrock Agent Core configuration
- MCP server implementation
- Deterministic deal scoring logic
- Brokered data access patterns
- Full Terraform IaC & Bitbucket CI/CD integration
- CloudWatch observability layer
Security & Responsible AI
Security & Data Governance
Zero direct AI access to data stores – all data brokered through defined service interfaces. AWS Secrets Manager with automated rotation. Field-level data classification controls. AWS Organizations SCPs constraining AI service behavior at the account boundary.
Operational Discipline
100% of agent activity surfaced via Amazon CloudWatch. ≥70% recommendation-to-SME alignment. Zero critical functional defects at program close. Deterministic rules and weights fully documented and adjustable by Airvoyant teams. Formal evidence package for go/no-go decision.
