AI-powered analysis of civil engineering services in fiber-optic network expansion

Project overview

• Client: OXG Glasfaser GmbH, a provider of telecommunications infrastructure and fiber-optic network expansion

• Project goal: Automate the partial measurement verification process for civil engineering services using AI

• Team: up to four TNG experts

• Approach: Proof of Concept starting in May 2025, full implementation starting in July 2025

• Key Achievements: Objective and comprehensive verification of civil engineering component measurements regarding route lengths and surface classifications, rather than random spot checks

The situation

OXG plans to connect approximately seven million households to a fiber-optic network in the coming years. To achieve this, tens of thousands of kilometers of fiber-optic cables must be laid.

As civil engineering work began, the verification of contractors' services became an increasingly important focus. Billing is based, among other factors, based on the length of route laid and the type of surface, such as tarmac or paving. Until now, OXG employees verified the distances listed on partial measurement reports on foot using a measuring wheel. However, such a process cannot scale to support large-scale expansion.

Our target

OXG requested us to examine whether the partial measurement verification process could be automated using artificial intelligence. The goal was to provide transparent and complete documentation of the construction work actually performed while significantly reducing the manual verification effort.

Our approach

As a first step, we demonstrated the technical feasibility through a proof of concept. Building on this, we began developing a working solution two months later.

To document civil engineering work, construction service providers create digital scans of the routes during construction. This image data forms the foundation of our solution. For image segmentation, we trained a neural network that automatically recognizes and classifies different road and sidewalk surfaces.

We use LabelMe as an annotation tool to generate the training data. Furthermore, we implement the training pipeline in Python using PyTorch and evaluate training progress with TensorBoard.

Additionally, we developed a web application that visualizes the scans on a map. To achieve this, we process the geospatial data using GDAL and provide the imagery as Cloud Optimized GeoTIFFs (COGs). Within the application, construction contractors can. Construction service providers can use it to:

• view the route of the excavated trench,

• accept the automatically determined route lengths, and

• manually correct surface classifications as needed, for example, if scans are missing, displaced, or obscured.

All adjustments are directly incorporated into the basis for subsequent invoice verification.

Technology-wise, we use Python on the backend and TypeScript, React, and OpenLayers on the frontend. The solution runs on AWS.

The result

With the new software, construction service providers can view and accept the calculated route alignments and route lengths directly when submitting their partial measurements. This provides OXG with a transparent and easily trackable basis for verifying the submitted services.

A key benefit here is the scalability of the process: Instead of time-consuming on-site inspections or random spot checks, it is now possible to conduct a comprehensive evaluation of the billed services.

Even after the successful go-live in May 2026, we continue to develop the application in collaboration with OXG. The focus is on expanding the web application to include drawing tools that allow missing route segments to be added directly, as well as analyzing trench depth and width for even more accurate billing.

As part of Big Techday 26, our client and we gave a talk on this project. You can watch the full recording here.