JanSahayak — When AI Becomes a Bridge Between Citizens and the Support They Deserve

The Gap Nobody Talks About

India has hundreds of welfare schemes. Scholarships for first-generation college students. Health coverage for families below the poverty line. Legal aid for those who can’t afford a lawyer. Food support for the elderly. Disability grants. Local municipality initiatives. They exist. They are funded. They are waiting. And yet — millions of eligible citizens never access them.

Not because they don’t need them. Not because they didn’t try. But because the information is locked inside dense government PDFs, written in bureaucratic language that even educated citizens struggle to parse. Because the eligibility rules span multiple paragraphs with nested conditions. Because the application process involves checklists nobody hands you. The gap isn’t the scheme. The gap is access to understanding. That’s the problem JanSahayak was built to solve.

What Is JanSahayak?

JanSahayak (Hindi: जनसहायक — “Helper of the People”) is an AI-powered Community Benefits Navigator designed to simplify access to welfare schemes and public support programs. Users can upload a wide range of documents, including:

👉 Government scheme PDFs

👉 NGO eligibility guides

👉 Scholarship circulars

👉 Policy documents

With this, JanSahayak becomes an intelligent assistant capable of answering questions in simple, actionable language. Instead of navigating complex documents manually, citizens receive:

👉 Eligibility insights

👉 Benefit explanations

👉 Required documents

👉 Step-by-step guidance

Fig1: From Question to Clarity in Seconds

How It Works

JanSahayak is built on a RAG (Retrieval-Augmented Generation) architecture powered at its core by NVIDIA Nemotron 3 Super -a model purpose-built for long-context reasoning, multi-step planning, and tool use. This makes it uniquely suited for the task: policy documents are long, eligibility rules are nested, and answers need to be grounded in actual source text, not hallucinated.

System Architecture

JanSahayak follows a Retrieval-Augmented Generation (RAG) architecture designed to process complex welfare and policy documents and convert them into accurate, user-friendly guidance.

The system combines document ingestion, semantic retrieval, vector search, and grounded AI-generated responses to ensure reliability and contextual understanding.

1. Document Ingestion Layer

The ingestion pipeline handles uploaded government scheme PDFs, scholarship circulars, NGO documents, and policy guidelines.

These documents are often:

• unstructured,
• lengthy,
• inconsistently formatted,
• and difficult to parse manually.

To address this, the system extracts:

• eligibility criteria,
• benefit details,
• deadlines,
• required documents,
• and application procedures.

This layer was designed to transform raw policy documents into machine-readable structured information for downstream processing.

2. NVIDIA Nemotron 3 Super

At the core of JanSahayak is NVIDIA Nemotron 3 Super, chosen for its:

• long-context reasoning,
• multi-step understanding,
• and ability to process complex textual workflows.

Government policies frequently contain nested eligibility conditions spread across multiple sections and pages. Nemotron enables the system to understand these relationships while generating grounded and context-aware responses.

The model is used for:

• document understanding,
• contextual reasoning,
• response generation,
• and extracting actionable insights from uploaded documents.

3. MongoDB

MongoDB is used as the primary database for storing:

• processed document metadata,
• extracted policy information,
• user queries,
• and application-related structured data.

MongoDB was selected because of its flexibility in handling semi-structured and evolving document schemas, which is common in policy-based datasets.

It helps maintain scalable and organized storage for both document content and application workflow data.

4. Milvus Vector Database

Milvus is used as the vector database for semantic search and retrieval.

After processing documents, embeddings are generated and stored in Milvus. These embeddings allow the system to perform similarity searches across policy content instead of relying only on keyword matching.

Milvus was chosen because it supports:

• high-performance vector search,
• scalable indexing,
• and efficient retrieval for large document collections.

This enables JanSahayak to quickly identify the most relevant policy sections when users ask questions.

5. Retrieval-Augmented Generation (RAG) Pipeline

The RAG pipeline ensures that all generated responses are grounded in actual document content.

When a user submits a query:

1. The system converts the query into embeddings.
2. Milvus retrieves the most relevant document chunks.
3. Retrieved context is passed to NVIDIA Nemotron.
4. Nemotron generates a contextual and human-readable response.

This architecture was chosen to minimize hallucinations and improve factual reliability, which is especially important in welfare and public-service applications.

6. Streamlit Interface

The user-facing interface is built using Streamlit.

Streamlit enables users to:

• upload policy documents,
• interact with the chatbot,
• view extracted information,
• and receive structured guidance.

It was selected because of its rapid development capabilities and ability to quickly prototype AI-driven workflows with an interactive UI.

7. End-to-End Workflow

The overall workflow of JanSahayak can be summarized as follows:

1. Users upload policy or welfare-related documents.
2. The ingestion layer extracts and processes content.
3. Embeddings are generated and stored in Milvus.
4. Structured data is stored in MongoDB.
5. User queries trigger semantic retrieval.
6. Relevant context is sent to NVIDIA Nemotron.
7. The system generates grounded, actionable responses for the user.

This architecture enables JanSahayak to transform complex government documents into accessible, understandable, and actionable guidance for citizens.

Fig 2 :Flow diagram

Fig 3 :Architecture

The Tech Stack

👉 NVIDIA Nemotron → AI extraction & reasoning 👉 MongoDB → Data storage 👉 Milvus → Vector search 👉 Sentence Transformers → Embeddings 👉 Python → Backend logic 👉 Streamlit → UI

Fig 4 : From raw policy documents to actionable answers — the technology stack powering JanSahayak.

A Real Interaction

Here’s what a conversation with JanSahayak looks like:

Fig 5 :Sample Interaction with JanSahayak

Snippet of Output

This snippet demonstrates how JanSahayak transforms complex policy documents into clear, structured, and actionable guidance. Instead of navigating lengthy PDFs, users receive:

• Eligibility insights
• Benefits breakdown
• Required documents
• Step-by-step application instructions

All responses are grounded in actual document content, ensuring accuracy and reliability.

Fig 6: JanSahayak interface showcasing extracted eligibility and benefits information from policy documents in a structured, user-friendly format.

Why This Matters to Us ?

Our team comes from communities where this gap is visible and real. We’ve seen family members miss scholarship deadlines because the circular was in English and the family wasn’t. We’ve seen elderly relatives lose out on pension schemes because nobody explained the eligibility criteria clearly. JanSahayak is personal. It’s not a demo project. It’s something we’d actually give to our families. The NVIDIA Nemotron 3 Super Build Contest gave us the compute, the model, and the community to build it fast. The Docker Bangalore and Collabnix community gave us the energy to ship.

Challenges We Faced Building this wasn’t without friction:

Unstructured PDFs-Government documents are a formatting nightmare. Tables split across pages. Eligibility rules written as paragraphs rather than structured data. We built a multi-pass extraction pipeline using Nemotron’s long-context reasoning to handle this.

Accuracy -In the welfare domain, a wrong answer isn’t just unhelpful ,it’s harmful. We grounded every response strictly in source documents using RAG, and added confidence indicators when retrieved context was weak.

System Integration -Getting MongoDB and Milvus to play nicely with a real-time Streamlit interface required careful API design and async handling.

Large Documents -Some scheme circulars run to 80+ pages. We implemented overlapping chunking with context preservation so no critical eligibility clause gets lost at a page boundary.

What’s Next We’re not done

JanSahayak’s roadmap includes: —

Eligibility Checker — A structured form that walks users through qualification criteria before they read a single line of policy —

Checklist Tracker — Persistent tracking of application progress, document collection, and deadlines —

Multilingual Support — Hindi, Tamil, Bengali, Kannada, and more because “plain language” means nothing if it’s in the wrong language .

Mobile App — A lightweight Android and iOS app for citizens who don’t have laptops .

Proactive Alerts — Push notifications when a scheme the user might qualify for opens applications .

The long-term vision is a unified, AI-powered welfare navigator covering every central and state scheme in India -a single place where any citizen can ask “what support am I eligible for?” and get a real, actionable answer.

Open Source Code

The complete implementation is available on GitHub.

GitHub — DeepMathukiya/community_navigator · GitHub

PPT: https://docs.google.com/presentation/d/1M5_I6Bd8jEC2hvf8b3AcArE1nfqtVaha/edit?usp=drive_link&ouid=114053297310419900461&rtpof=true&sd=true

YouTube :

https://youtu.be/buFs5PGKhss?feature=shared

Conclusion

JanSahayak is more than just an AI chatbot -it is a step toward making public services truly accessible. It is an attempt to make public support systems truly accessible. By transforming complex policy documents into clear and actionable guidance, JanSahayak helps bridge the gap between information availability and real-world access.

We built JanSahayak as part of the NVIDIA Nemotron 3 Super Build Contest hosted by Docker Bangalore × Collabnix × NVIDIA AI. A huge shoutout to the Bengaluru tech meetup community — the Collabnix community, Docker Bangalore organizers, and every contributor who showed up, gave feedback and kept the energy alive.

NVIDIA AI #Nemotron #NemotronBuildContest #JanSahayak #CivicTech #DockerBangalore #Collabnix #RAG #Milvus #BuildWithAI #BengaluruTech #CommunityBenefitsNavigator NVIDIA Web Development Docker Ajeet Singh Raina

👨‍💻 Built by Team JanSahayak:

Dsmathukiya | Himsikha | Ramakrishna Kancherla | Sunil Kumar