On-device AI · LinkedIn adaptation
Can a small AI model be useful offline?
A Ganeshotsav model for aartis and shlokas became a practical test of what fits on a phone.
Ravi’s Ganeshotsav project started with a constraint: the assistant had to run on a phone after the model was downloaded. He chose a narrow set of tasks, including reciting familiar aartis and shlokas and answering simple festival questions in Marathi, Hindi, or English.
Evaluate what people actually ask
The first small model could recite passages when prompted in English and still struggle when asked for the same thing in Marathi. The project expanded the training examples and added held-out tests for native-language prompts. Exact recitation gave the project a clear way to tell whether changes helped.
Design for the phone
The larger model ran too slowly for a comfortable phone experience. The project moved to a smaller Gemma 3 model and packaged it for Google AI Edge Gallery. The release file is about 1 GB. The model card gives the setup steps and documents the evaluation.
Know when to check another source
The model card reports weak results on ritual steps and story variants. Verify those answers with a trusted source and check a panchang for timing. If you’re building an offline assistant of your own, start with a task you can check directly, test it on the device people will use, and keep a record of where it fails.