Padmanabh Kulkarni
← BACK TO PROJECTS
Mobile DevelopmentLLMGemma3n

Nimittam: AI Companion for Mass Gatherings

Nimittam is an offline-first AI assistant built on Google's Gemma 3n, designed to transform any Android phone into an intelligent guide for large-scale events like the Kumbh Mela and Hajj. It provides real-time multilingual translation, visual landmark navigation, and crowd safety alerts, entirely on-device, no internet required.

CATEGORY: ANDROID DEVELOPMENT
Nimittam: AI Companion for Mass Gatherings

The Problem

The Maha Kumbh Mela 2025 hosted over 400 million visitors across 4,000 hectares, which is the largest human gathering on Earth. Events at this scale bring devastating challenges that existing technology fails to fully solve:

  • Lost and separated persons: tens of thousands of missing person cases were registered at Kumbh 2025 alone

  • Stampede and crowd surge risks: millions packed into dense zones with limited real-time rerouting

  • Language barriers: pilgrims from across India and abroad speaking dozens of languages and dialects

  • Network failure: cellular networks collapse under the load of millions of simultaneous users

  • Navigation confusion: first-time visitors overwhelmed by a temporary city the size of a small town

Any real solution must work offline, in any language, on a basic Android phone.

The Solution

Nimittam turns a smartphone into a context-aware, intelligent pilgrim companion. A user can speak or type in their native language, point their camera at a signboard, landmark, or crowd, and receive instant, meaningful guidance which can all be processed locally on the device using Gemma 3n.

Key Features

  • Offline Multilingual Voice Assistant: understands and responds in Hindi, Tamil, Bengali, Marathi, Gujarati, and 13+ other Indian languages with no internet required

  • Visual Navigation: point the camera at any tent, signboard, or landmark; the app recognizes it and provides step-by-step directions

  • Crowd Safety Alerts: proactively warns users when sectors are overcrowded or temporarily restricted, with updates sideloaded via Bluetooth or minimal Wi-Fi sync

  • Live Translation: translates spoken language, typed text, and printed signage captured via camera using on-device OCR

  • Lost Person Assistance: helps locate and reunite separated individuals through privacy-preserving, anonymous communication channels

  • Emergency Mode: displays emergency messages in the user's local language that can be shown to bystanders, usable without any connectivity

  • Accessibility First: large buttons, icon-based guidance, and offline help tips designed for elderly and digitally unfamiliar users

How It Works

The entire AI stack runs locally on the user's device. On first launch, the user downloads a Gemma 3n compatible model file once from a trusted source like Kaggle. From that point on, no cloud calls are ever made for inference. When a user sends a text query or captures an image, the Kotlin app composes a multimodal prompt combining the image and text, passes it to the on-device Gemma 3n model, and renders the response instantly using native Android components.

System Architecture

The core architecture is built around three layers:

Gemma3n Architecture.png

  1. Kotlin Application Layer: handles user interaction, camera input, permissions, and input/output coordination

  2. On-Device AI Layer: Gemma 3n (2B or 4B variant) running via MediaPipe API, handling language understanding, image description, translation, and navigation responses

  3. Local Storage Layer: stores the downloaded model, cached map data, and landmark recognition datasets entirely on-device

Technologies Used

  • AI Model: Gemma 3n (2B / 4B variants), quantized via INT8 / Float16 for mobile deployment

  • Runtime: MediaPipe API for on-device multimodal inference

  • Platform: Android (Kotlin), minimum Android 12 (API 31)

  • Vision: Multimodal image + text encoding pipeline using vision-language embeddings

  • Target Devices: Mid-tier and low-end Android phones (Snapdragon 720G, 8GB RAM and below)

1-combine.jpg

Real-World Use Cases

Nimittam is designed for any scenario where millions gather without reliable infrastructure:

  • Kumbh Mela and Hajj: offline navigation, multilingual assistance, and crowd safety for pilgrims

  • Stadiums and Concerts: offline seat-to-exit mapping and crowd egress routing

  • Refugee Camps and Disaster Shelters: APK sideloaded via Bluetooth, running on solar-charged phones for registration and missing person tracing

  • Pilgrimage Trains and Rural Fairs: OCR-based station board translation for zero-connectivity travel

  • Smart City Crowd Management: GDPR-compliant people counting with sub-500ms latency on local edge hardware

TEAM

Built for the Google: The Gemma 3n Impact Challenge by Padmanabh Kulkarni, Rithik Purohit, Mahesh Lambe, Krishna Shingan, and Ravikant Khamitkar.

links

Download APK: https://github.com/Nimittam/nimittam-app/releases/download/V1.2/nimittam_V1.2.apk

Kaggle Write-up: https://www.kaggle.com/competitions/google-gemma-3n-hackathon/writeups/nimittamai

YouTube Video: https://www.youtube.com/watch?v=aEIQ1ODBoyQ&t=1s

GitHub Repository:https://github.com/Nimittam/nimittam-app