MININDU
BIMSARA

A six-node LangGraph DAG with shared immutable state for full audit traceability. Sensor Analyst, AI Validation Engineer, Diagnostic Classifier, Knowledge Retriever, Execution Strategist, and Safety Critic operate sequentially — with a bounded two-retry refinement loop enforcing LOTO/PPE compliance before any procedure is surfaced to the operator.

Document ingestion through four stages: YOLOv8/DocLayNet layout detection → GPT-4o Vision captioning of engineering schematics into domain prose → 500-token sliding-window chunking → indexing into PostgreSQL/pgvector. A unified 1536-dimensional embedding space enables cross-modal retrieval of text, tables, and diagrams without auxiliary visual indices.

Resolved incidents traverse a five-gate quality pipeline before archival: Safety Critic approval → per-step AI verification → operator resolution confirmation → GPT-4o summarisation → vectorisation into pgvector interaction memory. Operator field expertise is captured into queryable organisational memory without model retraining.