Prevent Construction Accidents with AI Safety Analysis

Construction firms already utilize project management software like Procore or Autodesk Build for their daily safety logging and record-keeping. These platforms excel at storing vast amounts of data, including site superintendents' daily reports and associated photos. However, their core functionality is data capture and storage, not proactive risk analysis or safety compliance tracking as required by various regulations. A safety manager's critical role still involves manually reviewing every report, sifting through pages of text, and visually inspecting dozens of photos daily to spot potential issues.

Consider the common scenario for a safety manager overseeing multiple active projects, perhaps 10-15 sites for a regional contractor. If each site superintendent submits a daily report with an average of 15-20 photos, this creates a daily workload of 150-300 photos and 10-15 detailed text reports to review. This manual data review, similar to estimators flipping through 50+ drawing pages per project for takeoffs, becomes a significant bottleneck. Managers spend hours each morning scanning this information, often looking for obvious problems. A recurring but subtle hazard, like specific tools improperly stored, frayed electrical cords appearing for three consecutive days at a particular site, or consistent minor PPE violations, is easily missed within this sheer volume of data. The critical information is logged, but without intelligent processing, it remains raw data, not actionable intelligence.

While some firms attempt to use business intelligence (BI) tools like Tableau or Power BI to visualize safety metrics, these tools cannot interpret the unstructured content within a photo or the narrative fields of a daily report. You can chart the number of reports filed, but you cannot easily chart the frequency of 'unsafe scaffolding' mentions or visually confirm if hard hats are consistently worn across all photos. The fundamental challenge is that traditional project management and BI systems are built for structured data and record-keeping. They lack the advanced AI capabilities to parse domain-specific, unstructured content like construction site photos and free-text safety observations. This absence creates a gap between data collection and true proactive risk mitigation.

Our engagement for a safety analysis system begins with a thorough data audit and discovery phase. Syntora would connect to your existing project management system via its API – typically Procore, Autodesk Build, or similar platforms – to pull 3-6 months of historical daily reports and associated photos. This audit is crucial for understanding your specific safety vocabulary and identifying the most common and reliably detectable hazards within your data, such as missing personal protective equipment (PPE), unguarded edges, or improper material stacking. This initial phase culminates in a brief report outlining the top 3-5 hazards that can be consistently detected, alongside a fixed-scope proposal for building your custom safety analysis system.

The technical approach for such a system typically involves orchestrating a combination of large language models (LLMs) and computer vision models. A Python service, often deployed on a scalable platform like AWS Lambda or Google Cloud Functions, would be triggered each time a new daily report is uploaded. A model like the Claude API or another capable large language model would parse the free-text content of the report, extracting specific keywords related to safety incidents, identifying locations, and classifying sentiment around safety observations. Simultaneously, a specialized vision model would analyze each uploaded photo to detect objects pertinent to safety, such as hard hats, safety vests, properly installed guardrails, or common hazards like debris and fall risks. The outputs from both the text and vision analysis are then correlated and stored in a structured database, such as Supabase, for easy querying and visualization.

The delivered system provides a concise, action-oriented risk dashboard, designed to complement your existing workflows rather than introduce another complex software suite. Each project site receives a daily risk score, for example, from 0 to 100. Any site exceeding a predefined threshold is flagged for immediate review, accompanied by a clear summary of the detected issues (e.g., 'Site 14: Risk Score 88 - Hard hats not detected in 4 of 18 photos; narrative mentions "unsecured materials" twice'). This dashboard can be configured to send daily summary emails directly to safety managers, enabling them to focus their attention efficiently on the highest-risk projects, rather than sifting through all daily reports. As part of our services, you receive all the custom-developed source code and a detailed runbook for system operation and maintenance, ensuring full ownership and transparency.