Agentic Generative AI for Context-Aware Outlier Removal and Historical Cost Optimization in Construction

Vamsi Sai Kalasapudi^1*Krishna Pranay Angara²Charles Tofferi³Naga Sai Aditya Varanasi⁴

• ¹University of North Florida, Jacksonville, United States
• ²SSS Soft Solutions, Orlando, United States
• ³Petticoat-Schmitt Civil Contractors, Jacksonville, United States
• ⁴Cognizant Technology Solutions, Jacksonville, United States

Digital timecards are widely used in construction to track labor hours, equipment usage, and productivity, yet they are prone to outliers caused by human error, inconsistent reporting, and interface complexity. These anomalies degrade data reliability, obstruct cost estimation, and limit the strategic use of historical performance records. Traditional outlier detection meth-ods, such as Z-score filtering and standard Isolation Forest, apply global thresholds that often fail to capture the heterogeneous and context-specific nature of construction data. This paper introduces a context-aware optimization approach that dynamically tunes Isolation Forest contamination thresholds by learning from estimating practices. Validation results demonstrate that, compared to Z-score filtering and standard Isolation Forest, the proposed method pro-duces tighter clustering of standard deviations across cost codes, eliminates extreme variance spikes, and better aligns actual productivity distributions with estimator expectations. The model effectively filters unreliable entries while preserving meaningful high-cost cases, thereby improving both interpretability and reliability of historical data. To support scalable use of these refined datasets, the authors developed a production-grade agentic AI workflow integrating es-timating and field management software with Google's Firebase and an OpenAI GPT-based assistant via OpenAPI specifications. This system allows estimating and project management teams to query their data conversationally, retrieving real-time productivity benchmarks, unit costs, and historical trends across jobs and cost codes. While the model currently functions as a post-correction mechanism rather than preventing errors at the source, it provides a scalable, automated alternative to spreadsheet-based workflows, enabling improved bidding, project planning, and business intelligence.