KumbhLabs Research

Building the intelligence layer
for the physical world.

We study how intelligent agents behave when they leave the browser and operate inside dense, multilingual, low-connectivity physical environments.

  • 01

    Agentic AI

    Personal agents that act on behalf of a person, with bounded autonomy.

  • 02

    Decentralized AI

    Coordination without a single centralized command system.

  • 03

    Edge Intelligence

    On-device retrieval and inference under weak connectivity.

  • 04

    Civic AI

    Agents as an interface between citizens, services and infrastructure.

  • 05

    Population Intelligence

    Understanding movement, demand and risk at mass-gathering scale.

  • 06

    Privacy & Trust

    Consent, minimal data and verifiable provenance by default.

Publications

Publication

Aug 1, 2026

KumbhDoot: A Scale-Ready, LLM-Bounded Architecture for Mass-Gathering Public-Service Assistants

A similarity-first architecture that bounds generative model usage behind semantic caching, approved information and intent routing - reducing cost and latency while keeping answers grounded.

Research

Jul 29, 2026

AI Agents for Kumbh Mela

A proposal for deploying agentic AI systems at the world's largest mass gathering, built around citizen-facing assistants, shared public infrastructure and scalable population intelligence.