Bengaluru may be India’s Silicon Valley, but when it comes to improving infrastructure, the city has largely drawn on conventional means. The BBMP Road Infrastructure Department coordinates with agencies and parasatal agencies to track and fix road issues.
But a simpler solution may be mounting a phone on a car dashboard to spot a pothole, record its location, and draft a complaint for the municipal official responsible for that stretch. The app powering this is only about 4 MB in size, detecting potholes in the cloud rather than on the edge.
It’s called Pothole Reporter and is the brainchild of Gaurav Sen, a Bengaluru-based software engineer and teacher who taught software engineering on YouTube for about nine years. His lessons have evolved from algorithms to cloud and distributed systems, and more recently, AI engineering.
Pothole Reporter is his first project directly addressing a civic problem.
“It was more like… what can we do about it? I don't want to write the complaints myself. It should be AI,” Sen tells AIM.
Complaint From a Camera
Sen initially built the tool for himself. He posted about it on Instagram, and he started getting some traction. Sen then opened the project on GitHub amid a huge influx of requests.
The app runs in the background while the car is in motion, sampling photographs through the phone’s camera. When it detects a pothole, it records the location and uses that information to identify the relevant municipal authority.
The system can also look through public tender records and match a location to the roadwork or contractor associated with it. Sen says the complaint can include details such as the pothole's size, location, and potential tender information before being routed to the appropriate official.
A person still has to review the findings and decide whether to go ahead with the complaint. The system does not automatically submit a complaint.
Government Data is the Harder Problem
Detecting a pothole is only one part of the equation. Matching it to the right authority depends on government data, which municipal bodies can publish in different formats.
“The data is actually out there. What is weird is that the data is published by different municipal corporations in different ways,” Sen explains.
He admits that building the same system several years ago would have required much more manual work to understand and connect those different data sources. AI now helps process inconsistent formats and deal with malformed data without requiring developers to manually account for every possible variation.
The app currently relies on OpenAI's vision capabilities for pothole detection. Expanding it to another city involves obtaining that city's tender information and connecting it to the location data.
“The reason it's easy to expand to pan India is that most of the intelligence is on OpenAI detecting a pothole,” Sen adds. At present, the app has been tested in some states and written test cases for others.
High Bar for Confidence
Sen has deliberately made the detection threshold conservative. The app flags a pothole only when the model is highly confident, leaving the final decision to the user.
“I have seen only one photo amongst all the photos that the app has taken that did not have a pothole but was still detected as a pothole,” he says.
However, on the contrary, the app also misses detecting some real potholes. Sen believes this happens more often than false positives.
He has also tested the system in difficult conditions. The app can detect potholes submerged in rainwater, while nighttime detection works when street lights provide enough visibility.
The app can currently work with a phone camera. Sen said a dashcam could also provide the required video stream, although that is not yet integrated.
Accuracy Comes at a Cost
Sen says processing an image through the current OpenAI setup costs him around half a rupee for every image. “If you are taking 400 photos in a day, then that's like ₹200,” he notes.
That balloons as the scale and reach of the project grows. In the long run, Sen does not see a cloud model processing every image for any state- or city-wide deployment. He is considering a smaller open-source model such as YOLO, which can be fine-tuned for a specific task.
A smaller model would also make the system cheaper and easier to scale if a government authority adopts it.
Moreover, if several people drive past the same pothole, the system could receive multiple reports about the same road defect. Sen says a central database would eventually be needed so the system could identify those duplicates.
That raises a question about who should control the database.
“Right now, it's on the user's phone, and OpenAI does get to see them,” he adds. OpenAI's contract states that it will not use the data to train its models.
If the system is centralised, the data then should ideally sit with the municipal corporation. If a private intermediary operates the system, that intermediary would have to take responsibility for storing it.
Government Interest
The project is attracting significant interest from governing bodies in Andhra Pradesh, Maharashtra, and Karnataka.
Sen has already had meetings with the Bengaluru Development Authority and Ramalinga Reddy, Minister for Transport and Chairman of Karnataka State Road Transport Corporation, Karnataka government.
“We are looking to get API access to the current Sahaaya grievance redressal system. For this, we need permission from the Greater Bengaluru Authority to whitelist our application. Once that happens, app complaints will go directly into their system and follow the standard resolution process,” Sen notes.
Pothole Reporter has received significant traction on social media. However, Sen believes the attention around the project does not mean much unless it produces a measurable result on the ground.
“People are upset about potholes, not happy with what I have created,” he remarks. He believes the real success would be being able to show that the system actually contributed to potholes being fixed.
The harder question is what happens after the AI identifies a pothole.
Sen has sent around multiple emails to the BDA but has not yet seen action on those complaints. To effect action, integration with a government system then becomes more important.
“One of the things I can do is I can give it to the government for free,” Sen shares.
What Lies Ahead
For now, teaching remains his primary work. His company, Interview Ready, helps software engineers prepare for technical interviews, while his AI courses focus on building production systems. Sen said about 500 students have attended his live AI cohorts, while his Interview Ready course has around 30,000 students and his YouTube channel has around 750,000 subscribers.
The pothole project has also changed the kinds of problems Sen wants to explore.
He describes another potential use of AI involving domestic workers who struggle to navigate government paperwork and financial services. His idea is to create a WhatsApp-based assistant that could tell users which documents they need, which office to visit, and how to escalate a problem.
“Just trying to use AI to tell them exactly which documents to carry, exactly which office to go to, how to escalate, having the ability to draft an email and send it would be great,” he shares.
For Sen, the broader question is whether AI can make existing systems easier for people to navigate.
“I think AI is going to bring a lot more transparency,” he asserts. “By equality, I mean at least the basics, the rights that we have, that we deserve, are going to be focused on a lot more.”
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.