Helmet

A Multi-Agent System for Intelligent Construction Site Management

Construction delays represent a critical industry challenge, with 98% of North American construction projects experiencing delays, extending project duration by an average of 37%. For a typical $750,000 build, each day of delay costs approximately $670, while larger projects face exponentially higher losses. This paper presents Helmet, a novel agentic architecture leveraging LangGraph and TiDB's vector capabilities to enable real-time construction site optimization.

Scenario

Consider this scenario: Two cranes in the South section await three trucks delivering wood pallets at 10:30 AM. Workers scheduled for 11:00 AM crane operations discover that yesterday's waste container blocks the site entrance. The traditional response involves hierarchical communication: worker → team leader → site manager → waste management company calls → 1.5-hour resolution time. This single obstruction cascades into delayed deliveries, idle cranes, reassigned workers, and potential multi-day schedule disruptions.

Our insight was simple: What if construction sites possessed reflexion, an intelligent system capable of preventing cascade failures through real-time optimization? Helmet embodies this vision through an agentic architecture that transforms construction sites into self-organizing, continuously adaptive systems. By leveraging the voices of workers on the field, their real-time observations, concerns and insights, we feed this critical human intelligence to AI agents that can instantly reorganize the worksite based on what workers are reporting and provide a clear action plan to the construction site supervisor. This creates a dynamic feedback loop where field expertise directly drives intelligent orchestration, automated resolutions completed in ~140/350 seconds. This enable the site to adapt and optimize in real-time as conditions change and challenges emerge.

From Monolithic to Agentic Architecture

Our initial approach followed conventional patterns: connecting a single Large Language Model (LLM) to task management APIs through multiple tools. This monolithic design quickly revealed fundamental scalability limitations:

Context Explosion: Providing multiple tools to a single model resulted in contexts exceeding manageable limits
Execution Latency: Single-agent operations required 400+ seconds for complex scenarios
Poor Specialization: General-purpose agents lacked domain-specific optimization capabilities

The transition to agentic architecture proved transformative. By decomposing functionality into specialized agents Planning, Conflict Detection, Notification, and Supervision we achieved:

Reduced Execution Time: From 400+ seconds to ~140/350 seconds for complex problems
Improved Specialization: Each agent optimized for specific construction domain tasks
Enhanced Scalability: Distributed processing enabling parallel task execution

Database Optimization

A critical learning involved our SQL Agent replacement. Initially, we deployed a general-purpose SQL Agent that dynamically generated database queries. This approach introduced significant overhead:

Schema Relearning: The agent repeatedly analyzed our database schema before query generation
Iterative Refinement: Multiple query attempts were required to retrieve specific information
~200-second Execution Delays: Each database interaction consumed substantial time

Our solution involved replacing the SQL Agent with a custom Model Context Protocol (MCP) server featuring 15 specialized database tools with strategic filters. This architectural change delivered:

Sub-second Tool Execution: Database operations now complete in under 1 second
Elimination of Schema Analysis: Pre-built tools encode domain knowledge

LangGraph Integration

LangGraph proved exceptionally well-suited for construction domain coordination. The framework's StateGraph paradigm naturally mapped to construction workflow patterns, enabling intuitive agent communication. The comprehensive documentation facilitated rapid implementation without significant learning curve overhead.

System Architecture

Helmet - Agentic

Helmet's core innovation lies in its specialized agent network orchestrated through LangGraph's StateGraph:

Supervisor Agent: Serves as the intelligent orchestrator, analyzing incoming requests and routing them through the specialized agent network based on request complexity and domain requirements.
Planning Agent: Functions as the strategic brain, executing direct SQL queries against TiDB to validate constraints, detect conflicts through our specialized Conflict Agent, and generate comprehensive resolution plans with alternative scenarios.
Notifier Agent: Ensures communication completeness by creating structured notifications with executable actions, maintaining clear audit trails, and providing one-click approval interfaces for construction supervisors.
Executor Agent: Handles atomic application of approved changes, ensuring data consistency and safe execution of complex multi-task reorganizations.

Helmet leverages TiDB's vector search capabilities to solve the semantic matching problem inherent in construction communication. When workers report issues using natural language "the electrical thing in the office is broken" our vector search automatically identifies relevant tasks and qualified personnel.

Task Vectorization: Our function creates semantic representations incorporating:

Task title and detailed description
Location data (room, building section, zone type)
Trade categories and skill requirements
Required materials and equipment specifications
Safety requirements and additional notes

User Vectorization: Our function encodes worker capabilities:

Complete identity and role information
Primary and secondary skill sets
Trade category expertise and experience levels
Project experience and skill level mappings

This approach eliminates the need to retrieve entire task databases, instead returning the 3-5 most semantically relevant matches, dramatically improving response times while maintaining accuracy. This also allows us to find the best worker for a task based on the skills required.

Construction Site Data Model

Helmet - Hoverville Retail Park

Our database schema reflects the complex interdependencies inherent in construction projects. The core tables Users, Tasks, Notifications, Messages, UsersVector, TasksVector capture both the hierarchical nature of construction organization and the intricate web of task dependencies.

The Tasks table incorporates sophisticated constraint modeling:

Location Hierarchy: room → floor → building_section → zone_type
Skill Requirements: trade_category → skill_requirements → required_materials → required_equipment
Temporal Dependencies: start_date → due_date → dependencies → blocks_tasks
Resource Allocation: assigned_workers → required_worker_count → safety_requirements

This granular modeling enables our Planning Agent to understand complex constraint interactions and generate feasible reorganization scenarios that respect both physical limitations and worker capabilities.

Performance Validation

We validated Helmet using the blocked entrance cascade failure scenario.

Problem: Waste container blocking site entrance, threatening delivery schedules and crane operations.
Traditional Resolution: 1.5+ hours of hierarchical communication and manual reorganization.
Helmet Resolution: ~140/350 seconds end-to-end, delivering alternative routes, crane reassignments, worker reallocation, and timeline adjustments.

Key Metrics: ~140/350-second average resolution for complex problems using 15 specialized MCP tools, with sub-second tool execution. Current challenge: occasional 40-second inter-agent communication delays.

Technical Challenge

Communication Optimization: Addressing LLM response interpretation delays between agents through structured output schemas.
Data Synchronization: Extending from static planning to real-time dynamic optimization.
Safety Validation: Implementing fail-safe mechanisms for critical construction operations.

Conclusion

The construction industry stands at an inflection point. With 98% of projects experiencing delays and cascading cost implications affecting all stakeholders, the need for intelligent coordination systems has never been more urgent. By reducing cascade failure resolution from hours to minutes, we've proven that construction sites can achieve unprecedented operational efficiency through intelligent automation.

Our technical journey from monolithic LLM architectures to specialized agent networks, from general SQL agents to custom MCP servers illustrates the importance of domain-specific optimization in complex industrial applications.

Our Open Source commitment ensures that the broader construction technology community can build upon these foundations. As we continue developing Helmet, we remain focused on our core vision: construction sites that never stop building, powered by every worker's voice transformed into instant intelligent action.

Every second of construction downtime represents money lost forever. We're ensuring those seconds become opportunities for optimization instead of sources of frustration.

SOURCE

Association of Professional Builders (2019) Alice Technologies (2024) Trangistics (2025) Cornerstone Projects (2022) Long International (2024) LogiNets (2025) Propeller Aero (2020) SmartPM (2025) Tungsten Capital (2019)