Struence

Real projects, measurable results.

Explore how we apply AI and machine learning to solve real-world construction engineering challenges.

VISIONRoad InfrastructurePeer-reviewed

YOLOv8 Asphalt Damage Detection

Manual visual inspection of road surfaces is time-consuming, subjective, and cannot scale to cover extensive highway networks. Different inspectors classify the same damage differently, and reactive maintenance leads to higher repair costs.

We developed a real-time computer vision system using YOLOv8 that detects and classifies 7 types of asphalt damage including cracks, potholes, net cracking, and patched areas. Combined with the Segment Anything Model (SAM) for precise area calculation, it enables automated PCI (Pavement Condition Index) estimation.

YOLOv8SAMPyTorchONNXTensorRTOpenCV
ASCE Journal of Transportation Engineering, Part B: Pavements · Vol. 152, Issue 1 · DOI 10.1061/JPEODX.PVENG-1880 (opens in a new tab)
mAP@50: 99.4%Detection Accuracy
mAP@50-95: 88.9%Overall Performance
7Damage Classes
VISIONQuality Control

AI-Powered Test Photo Classification

Industrial testing labs generate 200+ photos daily with meaningless filenames like "PHOTO-2025-07-08-12-11-51.jpg". Manual organization took 2+ hours daily, with frequent misfiling errors. Blurry photos were kept, and similar project names caused confusion between different clients.

We developed a hybrid OCR system combining a quality-filtering stage with a vision-language model for accurate label reading. The system automatically groups photos by capture time, identifies the best label photo in each group, extracts project and sample information, and organizes files into a hierarchical folder structure.

PythonOpenCVTesseract OCRVision-Language OCRTkinter
92%Classification Accuracy
85%Time Saved
$0.002Cost per Photo
INSIGHTProject Management

AI-Powered Email Assistant for Construction Projects

Construction project managers receive hundreds of emails daily containing critical information buried in attachments: schedules in Excel, reports in PDF, invoices and specifications. Finding specific information requires manually searching through emails and opening multiple attachments, wasting valuable time.

We built an intelligent email assistant that connects to Gmail, understands natural language questions, and searches through both email content and attachments (PDF, Excel, CSV). With dynamic query generation, users can ask questions like "What's the asphalt schedule for February?" and get instant answers with source citations.

React NativeNext.jsGmail APIRAG ArchitectureOn-device Inference
< 3 secQuery Response Time
PDF/Excel/CSVAttachment Support
Multi-languageQuery Support
INSIGHTDocument Intelligence

Cite-Rich Conversations with Document Collections

Teams working across large PDF collections (technical specifications, reports, contracts) cannot quickly find a specific fact, and when they do, they have no way to verify where the answer came from. Keyword search misses meaning and never points to the exact page.

INSIGHT turns a collection of PDFs into a conversation. Answers stream in with numbered citations, and clicking a citation jumps the built-in viewer straight to the cited page. A hybrid retrieval pipeline fuses semantic and keyword search so results match both meaning and exact terms, while a page-level OCR fallback keeps scanned, multilingual documents fully searchable with accurate page references.

Hybrid RetrievalVector SearchPage-level OCRStreaming CitationsSplit-pane Viewer
Click-to-sourceVerifiable Citations
HybridSemantic + Keyword Retrieval
MultilingualScanned-Doc OCR
FORENSICLitigation Support

Multi-Agent Legal Case Analysis

A single analyst reviewing complex litigation has to hold several legal lenses at once, criminal, civil, evidence, procedure and risk, across many documents. Angles get missed, and assembling a consistent, well-reasoned case report by hand is slow and error-prone.

FORENSIC ingests case documents, runs OCR and classifies each one, then convenes a council of five specialist agents that analyze the case in parallel from distinct legal perspectives. The lowest-confidence positions are challenged in structured debate rounds, and a coordinator synthesizes everything into one report with a tiered risk matrix and an aggregate confidence score. The full reasoning trail is preserved, not just the verdict.

Multi-Agent OrchestrationStructured DebateDocument OCRRisk SynthesisConfidence Scoring
5 agentsParallel Legal Council
Up to 5Debate Rounds
Risk matrixCoordinator Synthesis
LABAsphalt Materials

Marshall Stability Prediction System

Predicting asphalt mixture Marshall Stability values traditionally requires destructive testing, which is time-consuming and costly. Laboratories needed a faster, non-destructive method to predict stability values from mix design parameters while maintaining accuracy comparable to physical testing.

We developed an AI-powered prediction system using a 2-level stacking ensemble architecture combining RandomForest, XGBoost, and LightGBM. The system incorporates physics-informed feature engineering with domain-specific variables like Density_Ratio, Binder_Ratio, and Marshall_Quality_Index derived from asphalt engineering principles.

XGBoostLightGBMANFISOptunascikit-learn
R² = 0.791Prediction Accuracy
+6%vs. Literature Benchmarks
1716Samples Analyzed
LABEarthquake Engineering

Seismic Isolator Report Automation

Each engineer used different Excel and MATLAB templates for FPC test analysis, slowing down the workflow and causing inconsistent results across reports.

We built an end-to-end test automation system with automatic cycle detection algorithms, trapezoidal integration for hysteresis loop analysis, and standardized Excel report generation. The system supports multiple press data formats and produces EN 15129 / AASHTO compliant reports.

PythonNumPySciPyMATLABTkinter
95%+Time Savings
< 2 minAnalysis Time
100%Report Consistency