Machine learning for rural programmes

India's rural programmes generate millions of photos, records and geotags every month. Most of that data is collected to prove that work happened, and very little of it is ever checked. At the Ministry I built systems that check it automatically, at the scale the programmes actually run.

Work risk analysis for MGNREGA

Every MGNREGA worksite is recorded in the programme's MIS under a work category, with a photo as evidence. I built and deployed an image classification pipeline that predicts the category from each photo and compares it with the category recorded. It assigns a risk score whenever the two disagree, so officials can focus their verification on the works most likely to be wrong. It processes about 10 lakh images a month.

Road quality validation for PMGSY (eMARG)

Road inspections under the Pradhan Mantri Gram Sadak Yojana are scored and photographed through eMARG. I deployed a model that checks whether each road image and its assigned score are valid. NIC Bhopal uses it every day, sending about 5,000 images and getting back scores and validation details.

Finding unmapped habitations

A habitation that is not in the OMMAS database cannot be planned for a road. I combined Meta (Facebook) population data with records from NREGA, PMAY and SBM to find habitations missing from the database. Then I automated QGIS to produce A3 field-verification maps for about 1.6 lakh of them. Each map carries a QR code that opens the exact location in Google Maps, so field staff no longer type in coordinates by hand to reach the site.

What it taught me

The hard part was rarely the model. It was making the output useful to someone in a district office with a deadline and a weak connection, and earning the trust of the teams whose data we were checking. That lesson is the founding idea of SolveSystems Lab.