This App got us 1st Place at a Microsoft Hackathon

Leveraging the latest of Microsoft Azure Machine Learning to help solve a $6.4 billion infrastructure issue.

Brian Ruiz


This App got us 1st Place at a Microsoft Hackathon blog image

The Challenge

The main objective of the competition revolved around developing innovative solutions for improving the infrastructure of cities in the US. The 2019 Smart Infrastructure Hackathon was hosted at The Cannon (Houston), in partnership with Microsoft Azure. Thankfully, we had a team of really bright cloud solution architects for guidance.

Technology We’re Using

The technology we opted for the project revolved heavily around Python. For the phone data, we used a publicly available mobile app that allowed exporting the data to CSV files. We then leverage Python’s scripting ability to analyze the data from the phone, to use for the road pothole detection algorithm.

Python also allowed us to create data visualizations —matplotlib— of our data which we used to tweak the algorithm. All of this was tied together with the Microsoft Azure platform. From storing the data (CSV’s and images) in Azure, to using the Azure Machine Learning platform to train the model.

Tech Stack

Our Idea and Proposal

After juggling several ideas we settled on the issue of road potholes. We discovered that potholes cost American drivers $6.4 billion per year, which is in repair and insurance costs. A single pothole may cause are up to $300 per year per vehicle (2016), while the average price to repair a pothole is on average $30 — $50 per pothole. Thus the effort to aid this issue can be cost-effective.

  • Data collection and analysis
  • Pothole probability algorithm
  • Minimal Interface
  • Data visualization/mapping

Looking at our challenge from a high level we need the following components or stages, to be able to solve the problem. Also, we need to design a 2-layer confirmation of a road pothole. The first is the active dashboard camera, and the second uses the user’s phone telemetry data, such as GPS and accelerometer modules.

1. Analyzing mobile telemetry data

Every modern smartphone is equipped with very accurate telemetry systems. For example, the GPS and accelerometer, which we leveraged for location and to track any abrupt movement indicating a pothole on the road.

One of the difficulties with this approach was that logging every high acceleration movement may result in false positives of a road pothole. Below is also a time-series data visualization of the accelerometer data, generated with matplotlib.

import pandas as pd
import matplotlib.pyplot as plt

df = pd.read_csv('/home/br/Downloads/testdata.csv')
df['Time'] = pd.to_datetime(df.Time, format='%Y-%m-%d %H:%M:%S:%f')

columns = [


2. Live computer vision

This part was more difficult to implement, but this was our second layer for the prediction of potholes, implementation is done with some guidance from the Azure cloud architects.

We used Azure Logic Apps to import our road image frames to the cloud, and with OpenCV were able to register a model to start analyzing our content with the conveniently named Azure microservice Computer Vision.

Dashcam Footage
Dashcam Footage concept, showing the idea of live capturing and prediction of potholes.

3. The Interface

From the start, the idea was to keep the interface as intuitive as possible to minimize driving distractions. Therefore, we intended to create an interface that would show a dialog box only when our system predicted there to be a pothole, from there the user would confirm if so.

Unfortunately, we didn't have enough time to implement both the backend logic and the front-end interface. However, I created a mockup of the interface, which captures the essence of the idea, and it's a good example with an emphasis on a minimal user experience.

UI mockup
Demonstrating the idea of a simple pop-up modal to confirm a pothole. This is crucial to keep the driver's attention on the road.

4. Mapping pothole locations

Finally, we needed to plot our GPS data so the city could take action. The data we leveraged for this use-case included latitude and longitude coordinates, and date-time stamps. For geocoding, we used Folium, which is a Python library that allows us to plot data on a map. Location accuracy is important to pinpoint multiple occurrences of the same pothole from different users for clustering the locations.

import folium
import pandas as pd

SF_COORDINATES = (37.76, -122.45)
map = folium.Map(location=SF_COORDINATES)

for each in df.iterrows():
    location = [each[1]['Y'],each[1]['X']],
    clustered_marker = True

San Francisco
Map of San Francisco with potholes marked. Notice the clustering of entries, which is important for confirming a pothole.

The Final Result

By the time the hackathon countdown ended, we were able to design and implement a viable solution that would best present our initial idea --keep in mind this was a 24-hour challenge. Nevertheless, this was achieved thanks to our strong team collaboration and efforts, and some sprinkled guidance from the MS architecture team.

After presenting our MVP (minimum viable product), the judges called a draw for first place. Having won the competition granted us credits to some of Microsoft Azure cloud services like storage and processing. And, the company which hosted the competition granted us seats at their co-working space, which we could use to continue the development of our application and idea.

The team from UHD
The team from UHD


  • Internet of Things
  • Machine Learning
  • Hackathon


Questions or need more details? Ping me on the community chat or reach me at I'd be happy to connect!


Get personal updates and readings on topics like tech, design, productivity, programming, and more!

Join the 0 other readers.