From Manual Inspections to AI-Powered Defect Detection: How a ConTech Startup Cut Structural Assessment Time by 75%

• Client: A North American construction technology startup specializing in infrastructure inspection.
• Industry: Construction & infrastructure.
• Team size: 4 Cleveroad engineers (2 ML, 1 backend, 1 frontend) over 7 months.
• Core deliverable: AI-powered visual inspection platform processing drone-captured imagery to detect and classify structural defects in bridges, overpasses, and commercial buildings.

Structural inspections in commercial construction are slow, expensive, and dangerous. Certified inspectors rope-access or scaffold their way across bridge decks, retaining walls, and building facades — photographing defects by hand, then manually cataloging each crack, spall, or corrosion patch in spreadsheet-based reports. A single bridge inspection can take two to three weeks from field visit to final deliverable.

The client, a seven-person startup founded by two former civil engineers, had seen this firsthand across hundreds of projects. They tracked the cost: an average commercial bridge inspection ran $28,000–$45,000, with roughly 60% of that budget consumed by on-site labor and report generation. Missed defects — common at a 30–40% error rate for manual visual assessment, per the client's own field audits — led to rework, liability exposure, and in some cases delayed maintenance that compounded repair costs.

The founders had already partnered with a drone services company to capture high-resolution aerial imagery. Raw footage was not the bottleneck. Interpretation was. They needed a system that could ingest thousands of images per inspection, identify and classify defects by severity, and produce a structured report — without requiring a certified inspector to review every frame.

The client evaluated three vendors. Two proposed off-the-shelf object detection models fine-tuned on generic crack datasets. Cleveroad proposed something different: a custom training pipeline built on the client's own field data, with a classification taxonomy mapped to AASHTO bridge inspection standards. That specificity mattered. Generic crack detection returns bounding boxes. The client needed severity grades tied to engineering codes.

Cleveroad also offered a lean engagement model — a four-person team that could start within two weeks. The client did not have 18 months or a six-figure retainer to burn. Seven months, start to working prototype. That was the agreement.

Cleveroad designed and built an end-to-end inspection platform with three core layers: an image ingestion pipeline, a defect detection and classification engine, and a report generation interface.

Drone imagery arrives as geotagged JPEG and TIFF files, uploaded in bulk through a web portal or synced from DJI FlightHub. The pipeline normalizes resolution, corrects lens distortion, and stitches overlapping captures into composite orthomosaics where applicable. Each image is tagged with GPS coordinates, altitude, capture angle, and timestamp.

A two-stage model processes each image.

• Stage 1: a YOLOv8-based detector locates candidate defect regions.
• Stage 2: a ResNet-50 classifier assigns each region a defect type (crack, spall, delamination, corrosion, efflorescence) and a severity rating on a four-point scale aligned with the client's chosen inspection standard.

The model was trained on 14,000 annotated images from the client's historical inspection library, augmented with synthetic variations for underrepresented defect classes.

Detected defects populate a structured report template with annotated images, GPS-mapped locations, severity grades, and recommended action codes. Inspectors review the draft report in a web-based dashboard, confirm or override AI classifications, and export the final deliverable as a branded PDF. The review step typically takes 45 minutes to an hour — compared to the 8–12 days it previously required.

The platform runs on AWS, with compute-intensive inference workloads managed through EKS. Model training happens on EC2 P4d instances with NVIDIA A100 GPUs. Processed imagery and reports are stored in S3, with metadata indexed in PostgreSQL for search and retrieval.

The frontend is a React-based single-page application. Inspectors interact with a map-first interface — defects are plotted on a 2D projection of the inspected structure, color-coded by severity. Clicking a defect opens the source image with the detection bounding box overlaid.

Tech stack:

• Cloud: AWS (EKS, EC2, S3, CloudFront)
• ML frameworks: PyTorch, Ultralytics YOLOv8, torchvision (ResNet-50)
• Backend: Python (FastAPI), Celery for async task processing
• Frontend: React, Mapbox GL JS for spatial visualization
• Database: PostgreSQL with PostGIS extension
• Infrastructure: Terraform, Docker, GitHub Actions CI/CD
• Data labeling: Label Studio (self-hosted)

By the end of month seven, the platform was processing live inspection data from three pilot projects — two highway overpasses and one commercial parking structure.

Inspection turnaround dropped from 14–18 days (field visit to final report) to 3–4 days. The bottleneck shifted from image review to client scheduling. Defect detection accuracy reached 92.4% on the client's holdout test set, compared to the 60–70% consistency rate they had measured across manual inspections. False positive rates held below 6%, low enough that inspectors trusted the AI output as a first pass rather than a distraction.

Cost per inspection fell by an estimated 55–60%. The client's initial pricing model — offering AI-assisted inspections at 40% below the market rate for manual assessments — still preserved healthy margins.

The prototype also generated investor interest. Within six weeks of the first pilot results, the client closed a $2.8M seed round. The functioning platform — not a pitch deck — was the primary asset in those conversations.

Phase 2 is underway. The client is expanding the defect taxonomy to cover 14 additional defect types required for Department of Transportation compliance. Cleveroad is building a temporal comparison module that overlays inspections from different dates, flagging defect progression — a crack that grew 12mm in six months, a corrosion patch that doubled in area.

A mobile companion app is also in development, allowing field inspectors to capture supplementary images on-site and merge them into the AI pipeline in real time. The client expects to run 40+ inspections through the platform in its first full year of commercial operation.