Kumbh Sah'AI'yak: AI-powered chatbot that provides pilgrims with information and navigation support

The problem

The 2025 Mahakumbh Mela in Prayagraj, Uttar Pradesh, was one of the largest human gatherings ever recorded, with over 660 million visitors attending over 45 days. Visitors came from across India and beyond, speaking a wide range of languages.

Navigating an event of this scale presents practical challenges for visitors: finding accurate and timely information about rituals, travel routes, amenities, and safety; overcoming language barriers in a multilingual setting; and accessing reliable guidance on religiously sensitive practices. At the same time, the administering authorities needed to ensure that any public-facing information service would not interfere with ground operations or crowd management in the mela area.

The solution

To assist visitors navigating the Mahakumbh Mela, Samagra developed Kumbh Sah'AI'yak, an AI-powered chatbot available via WhatsApp and a web application in 11 Indian languages. The tool was developed in collaboration with the Prayagraj Mela Authority and UPDESCO, with technology partners including Bhashini (India's national language translation platform).

The chatbot provides pilgrims with information, navigation support, and a “memory” function, covering ten use cases across these three categories. Users can ask questions in text or voice about the event's history, rituals, travel options, accommodation, and on-site amenities, and receive responses in their chosen language.

The system uses a Retrieval-Augmented Generation (RAG) architecture, meaning it draws answers from a curated knowledge base rather than generating responses from general knowledge. It is powered by large language models and integrates Bhashini for machine translation and speech-to-text capabilities, allowing users to interact by voice in their own language.

To ensure interactions remain respectful and reliable, the system incorporates religious sensitivity safeguards and a three-step system for filtering inappropriate content, using Bhashini's language filter, Amazon Guardrails, and screening by the language model itself. The team also used an automated test bed for iterative testing and structured information as question-and-answer pairs to reduce the risk of the AI generating inaccurate responses.

The impact

1. Over 300,000 users across more than 100 countries

The chatbot was used by more than 300,000 unique visitors (3 lakh) before and during the mela, generating over 500,000 sessions. Users came from 102 countries. Travel and accommodation were the most common topics, and English, Hindi, and Gujarati were the most used languages.

2. 132,000 queries handled, with 70% answered successfully

The system processed over 132,000 text- and voice-based queries, with 70% answered successfully. Of the users who rated the service, 82% gave it 3 or higher on a 4-point scale, and 40% returned to use the chatbot more than once.

3. Content safeguards worked as intended

The three-step filtering system achieved a 100% success rate in not responding to profane or abusive content. No cases of profanity have been reported.

Lessons learned

• Reduce friction at the point of entry. The initial version required users to log in with a phone number, which created a significant barrier, as many visitors were hesitant to share personal information. Introducing a guest login option removed this friction and improved uptake. For any public-facing AI tool deployed at scale, minimising what users need to provide upfront matters.
• Design for how people actually communicate. The platform initially relied on predefined menus, but users expected to be able to ask questions in their own words. Pilgrims wanted to have open, unstructured conversations rather than navigate through set options. Meeting users where they are, rather than where the system assumes they will be, is important for adoption.
• Structured knowledge and iterative testing help control inaccuracy. Organising the underlying information as question-and-answer pairs and using an automated test bed for repeated testing helped reduce the risk of the AI producing inaccurate or fabricated responses.
• Content safeguards need to be designed for context. Given the religious significance of the event, ensuring respectful and accurate responses was essential. The three-layer filtering approach proved effective at preventing inappropriate content from reaching users.

Launch year: 2024

While there were no regulatory boundaries during the development of Kumbh Sah'AI'yak, the fact that Kumbh is a mass-gathering event was taken into account to ensure that the chatbot does not provide any answers which have the potential to impact the ground operations or crowd management by the administration in the mela area.