AccessMitra AI

AI that helps differently-abled Indians find government schemes, legal rights, and job opportunities they deserve.

Comment

The Problem Nobody Talks About

There are 2.68 crore differently-abled individuals in India. That is not a statistic. That is 2.68 crore people, each with a name, a family, a dream, navigating a system that was never built with them in mind.

The government has schemes for them. The RPWD Act 2016 guarantees them rights. NHFDC offers them loans. Niramaya offers them health insurance. Scholarships exist for their children. Job reservations exist in their name.

But here is the cruel irony: almost none of them know.

Not because they do not care. Not because they do not try. But because the information is buried in bureaucratic PDFs, scattered across twelve different ministries, written in language that assumes you already have a lawyer, a government contact, or hours to spend navigating portals that barely work.

I kept thinking, what if someone just told them? What if there was one place, in plain language, that said: here is what you are entitled to, here is how to get it, here are the documents you need, here is the exact link to apply.

That thought became AccessMitra AI.

What Inspired This

The inspiration was not a research paper. It was a conversation.

A family member mentioned someone in our community, a young man with locomotor disability who had been unemployed for two years, not because jobs did not exist for him, but because nobody had told him about the 4% central government reservation in his name. Nobody had told him about the NHFDC loan he could use to start something of his own. Nobody had told him he could get a free tricycle through the ADIP scheme.

He had rights. He had schemes. He had opportunities.

He just did not know.

That felt wrong in a way I could not ignore. And when the AI Apocalypse hackathon theme came, exploring how AI can reshape society, solve real problems, and fight for people the system forgets, I knew exactly what I wanted to build.

What I Built

AccessMitra AI is an RAG-powered knowledge assistant built specifically for differently-abled individuals in India.

At its core, it is a system that ingests verified government documents, Ministry of Social Justice and Empowerment scheme guidelines, RPWD Act 2016 full text, NHFDC loan scheme details, National Trust scheme documents, National Scholarship Portal guidelines, embeds them into a vector database, and lets anyone ask questions in plain English and get grounded, sourced, accurate answers.

But beyond the AI chat, I wanted it to feel like a companion, not a search engine. So I built:

  • An eligibility finder that asks five simple questions and instantly maps you to every central scheme you qualify for
  • A scheme explorer with all ten major central schemes explained in plain language, not bureaucratic language
  • A rights section that breaks down the RPWD Act 2016 into what it actually means for you, and what to do if those rights are violated
  • A document checklist for each scheme, the exact papers you need, the exact steps to follow, and the exact portal to visit
  • A jobs section that surfaces the employment reservations and resources most people never find

Every answer cites its source. Every scheme links to the verified official portal. Nothing is made up.

How I Built It

The stack was chosen for speed and reliability:

  • FastAPI for the backend, deployed on Render
  • Groq with llama-3.3-70b-versatile for generation, fast, free, accurate
  • Pinecone for vector storage, the knowledge base lives here
  • HuggingFace Inference API for embeddings, lightweight, no local model needed
  • Next.js with a custom green and amber design system for the frontend, deployed on Vercel

The knowledge base was carefully curated from verified public sources only: Wikipedia disability articles, official MSJE scheme PDFs, RPWD Act 2016 text, and National Trust scheme documents. Every piece of data was reviewed before being embedded.

The system prompt was written with deliberate care, instructing the model to always use the term differently-abled, to never guess eligibility criteria, and to always cite its source.

The Challenges

The biggest challenge was not technical. It was restraint.

It is very easy to build a chatbot that sounds confident and helpful while quietly making things up. For most use cases, a hallucinated answer is annoying. For someone asking, am I eligible for this health insurance? What documents do I need to apply? A hallucinated answer can cost them months of waiting, wasted effort, or a rejected application.

So the hardest thing I built was the discipline in the system, the grounding, the source citations, the honest acknowledgment when the knowledge base does not have enough information.

On the technical side, Render's free tier has a 512MB memory limit, which meant sentence-transformers and PyTorch had to go. Switching to HuggingFace Inference API for embeddings solved the memory problem without sacrificing quality.

The ingestion pipeline also went through multiple iterations before settling on a clean, reliable approach, flat top-level execution rather than wrapped functions, to avoid silent Python crashes on Windows during development.

What I Learned

I learned that the gap between an existing scheme and a person benefiting from it is enormous, and that gap is mostly an information problem.

I learned that building for a specific, underserved user is more meaningful than building for everyone.

And I learned that AI, when built carefully and honestly, can genuinely fight for people the system forgets.

What AccessMitra AI Means

Mitra means companion in Hindi. Not a tool. Not a portal. A companion.

Someone who sits with you and says: here is what you are entitled to. Here is how to get it. You deserve this. Let me help you find it.

That is what I wanted to build. And I think, for the first time, I built something that actually matters.

Built With

Share this project:

Updates

Submission history