Weld AI - Autonomous Multi-Agent Radiography

📌 Overview In regulated industries like oil & gas, piping, power, and manufacturing, weld structural failures can lead to catastrophic accidents. Quality assurance relies on Non-Destructive Testing (NDT) radiography, a manual, slow, and error-prone process. The Weld NDT AI Inspector transforms this workflow into a highly reliable, autonomous, and auditable multi-agent pipeline. The system is architected using Hexagonal (Ports & Adapters) design principles to keep core business rules independent of ML frameworks and databases, running securely on Google Cloud Platform (GCP) and MongoDB Atlas. 👥 The Multi-Agent Collaboration Room Inside a distributed Band.ai room, four remote agents work in sequence to ensure safe and compliant decision-making: Weld Orchestrator Agent: Directs the room flow, contrast-enhances the raw scan with CLAHE, dispatches payloads, and logs tamper-evident audit events. Weld Vision Agent: Deploys a fine-tuned computer vision model to locate and classify physical defects (e.g. porosity, slag, cracks) on grey-scale film. Weld Compliance Agent: Leverages a rules engine and Gemini to cross-reference dimensions against regulatory standard tolerances (ASME B31.3, API 1104, AWS D1.1). Weld Review Agent (Track 3 Core): A strict safety auditor that enforces mandatory overrides (e.g., any crack = mandatory reject, regardless of size) and routes ambiguous cases to an ESCALATE status for human Level III inspection. 💾 Enterprise-Grade Resiliency Dual-Database Adapter: Saves inspection records to MongoDB Atlas with automatic fallback to local SQLite storage to prevent downtime in offline industrial fields. Inference Caching: Indexes scans via cryptographic SHA-256 hashes, returning cached defect coordinates on duplicates to prevent redundant GPU overhead. Vertex AI Integration: Authenticates securely using GCP service account IAM policies, bypassing the need for static API keys.