The Indian government provides programs for farmers, students, women, business owners and medical patients. It is difficult for you to locate a specific program when you lack the title, the requirements for participation or a starting location. We intended to simplify this initial action.
There is a multilingual assistant we created called GramSeva where you describe your circumstances in English, Hindi or Kannada - using a keyboard or your voice to find relevant government programs.
The development process was difficult and complex. We began with a concept plus a small amount of knowledge regarding AWS then we studied Lambda, API Gateway, DynamoDB, S3, Bedrock, Transcribe, Polly & Amplify.
One of our first major setbacks was discovering that the Bedrock model we had selected wasn't accessible to our account. After investigating and resolving several issues, we used Bedrock Mantle to make the AI pipeline operational. Along the way, we also changed an important part of our approach. To improve reliability, AI is responsible for understanding the user's situation, while a deterministic rules engine handles scheme matching. Missing information remains unknown rather than being guessed or fabricated.
We reviewed the dataset, rebuilt the interface, managed production deployments, fixed failed scripts, and dealt with expired API credentials. Eventually, we reached our goal: we could speak Kannada, Hindi, or English to the live application, have it interpret the user's situation, and receive potentially relevant schemes.
Is only a partial answer to the bigger problem of finding welfare from government. It doesn't (yet) serve as a system of eligibility checks. But it is a functioning first step in improving ease of finding schemes.
We began with a simple question: "Can we build this?"
And guess what, we did.
Loading article