Every year, lakhs of workers leave their homes in Bihar, West Bengal, Odisha, and Tamil Nadu and arrive in Bengaluru looking for a better life. They build the city’s skylines, clean its homes, and deliver its food — yet most of them have little to no awareness of the legal protections and welfare schemes available to them.
The Karnataka Minimum Wages Act guarantees workers a minimum floor on their wages. The Employees' Provident Funds and Miscellaneous Provisions Act, 1952 mandates employer contributions toward retirement savings through EPF. The Employees' State Insurance Corporation provides workers and their families with access to healthcare through the ESI scheme. The Building and Other Construction Workers Act enables construction workers to register for welfare benefits, including scholarships for their children, maternity assistance, pensions, and accident compensation.
These rights and schemes exist publicly on paper — published by the Karnataka Labour Department, the Ministry of Labour and Employment, Employees' Provident Fund Organisation, and Employees' State Insurance Corporation. However, the information is typically buried in English-language PDFs and fragmented government portals that assume literacy, internet familiarity, and bureaucratic knowledge. For a worker who speaks only Hindi, Bengali, Odia, or Tamil, these systems are effectively inaccessible.
To bridge this gap, we built the Karnataka Worker Rights Assistant.
The platform begins with the Elastic Open Crawler, which continuously indexes official government sources including:
- labour.karnataka.gov.in
- labour.gov.in
- epfo.gov.in
- esic.gov.in
- the Karnataka BOCW Welfare Board portal
The crawler ingests wage notifications, circulars, welfare scheme details, eligibility criteria, and registration procedures directly from official government pages. Every indexed document remains traceable to its original source URL, ensuring that every answer generated by the system can be independently verified.
The crawled documents are chunked and embedded using multilingual Jina v5 embeddings through Elasticsearch’s Elastic Inference Service. This enables a single semantic index to understand and retrieve relevant information across multiple languages. A question asked in Hindi retrieves the same relevant legal passages as the equivalent query in English.
When a worker submits a question, the system performs a hybrid semantic and keyword-based retrieval over the indexed corpus. The most relevant passages — along with their official source URLs — are then passed to Anthropic Claude running on Amazon Web Services Bedrock.
The system prompt enforces three strict rules:
- Respond entirely in the worker’s language.
- Cite an official government source URL for every factual claim.
- If the information is not present in the retrieved documents, explicitly say so and direct the worker to the Karnataka Labour Department helpline instead of fabricating an answer.
The complete agent runtime is hosted on AWS AgentCore, which provides managed serverless infrastructure for deploying, scaling, and monitoring the assistant.
Before any user query reaches the retrieval pipeline, it passes through a security layer that:
- Detects prompt injection attempts
- Scrubs personally identifiable information (PII)
- Removes sensitive data such as Aadhaar numbers, phone numbers, and bank details
This ensures that vulnerable workers can safely ask questions without accidentally exposing sensitive personal information.
Built With
- amazon-web-services
- elastic
- python
Log in or sign up for Devpost to join the conversation.