Multi-Agent Architecture

The 8-Agent Ensemble

Detailed breakdown of every specialized AI agent in the STRAND LangGraph ensemble.

Guardian Agent

Specification Compliance & OCR Ingestion

Performs OCR on engineering submittals, matches parameters against tender requirements using regex & NER, computes R0 contagion impact, and drafts formal RFIs.

Inputs

PDF cut-sheets, engineering submittals, vendor drawings

Outputs

Spec-DNA nodes, violation alerts, automated RFI drafts

Scheduler Agent

CPM Network & Delay Propagation

Builds a directed acyclic task graph (DAG), simulates Critical Path Method (CPM), calculates R0 infection scores, and recommends schedule compression buffers.

Inputs

Primavera P6 schedule CSV, active submittal deviations

Outputs

Critical path analysis, task delay probabilities, R0 risk ranking

Oracle Agent

Supply Chain Risk & Vendor Benchmarking

Monitors global logistics lanes, predicts port delays, and evaluates alternative suppliers across lead time, cost variance, and historical defect rates.

Inputs

Active equipment shipments, GPS coordinates, weather/geopolitical feeds

Outputs

GeoJSON transit maps, supplier risk scores, fallback vendor rankings

Brain Agent

Hybrid GraphRAG Intelligence

Fuses BM25 keyword matching with dense vector retrieval using Reciprocal Rank Fusion (RRF), enriched by 1-hop Neo4j BIM graph traversal.

Inputs

Natural language questions from engineers and project managers

Outputs

Grounded answers with exact document citations and 1-hop BIM relationships

Inspector Agent

Multimodal Voice Commissioning

Enables field engineers to speak checklist verifications on-site, transcribe voice notes into structured Non-Conformance Reports (NCRs), and sync offline.

Inputs

Field audio voice notes, photo evidence, checklist verifications

Outputs

Structured Commissioning NCRs, offline sync queue, As-Built Markdown

Planner Agent

Strategic Mitigation & HITL Gate

Synthesizes multi-agent findings into comprehensive recovery action plans. Enforces Human-in-the-Loop (HITL) approval gates before executing write operations.

Inputs

Multi-agent findings, open NCRs, schedule delay alerts

Outputs

Executive mitigation plans, pending HITL approval items

Judge Agent

Grounding Grader & Anti-Hallucination

Acts as an independent verifier that checks whether all claims made by other agents are factually grounded in raw ingested PDF chunks.

Inputs

Generated LLM answers, raw source document chunks

Outputs

Grounding confidence scores, compliance sign-offs

Dashboard Agent

Natural Cypher Widget Generator

Generates custom visual widgets on the fly by converting user prompts into safe read-only Cypher queries.

Inputs

User natural language requests ('Show me all open NCRs by vendor')

Outputs

AST-sanitized read-only Cypher queries, interactive dashboard widgets