Amanda Fitch used Google’s Antigravity agent, EmbeddingGemma 2 and Flutter to create Lark, a semantic search app for finding personal photos, recordings and documents.
Amanda Fitch has shown how a highly personal search tool can be built for a phone without writing code, at least in the first instance. In a Medium post, she described wanting a better way to search for material linked to a hamster called Stinky across photos, videos, journal entries and songs on her device, rather than relying on a basic file-title search. Her aim was a semantic search app that could understand themes, not just filenames.
To find a model that could run on a handset, she looked through Hugging Face and settled on Google’s EmbeddingGemma 2 after comparing it with other multimodal options. She described the model as small enough to fit on a smartphone, which made it a practical choice for local use. For the app framework, she chose Flutter, Google’s cross-platform toolkit, so the same codebase could support mobile and other devices.
Fitch then used Google’s Antigravity agent to generate an app called Lark, asking for a cheerful interface that could search photos, videos, documents and voice recordings. Within about 45 minutes, she had an initial version. But the first build did not truly connect to a model; it only simulated search results with mock data. After several hours of refinement, she said the agent was instructed to remove the placeholder behaviour and link the app properly to the embedding model.
She later asked the agent to shift the app away from on-device processing and towards Google’s Model Garden, which she said could free up storage on the phone and make the tool usable across devices. Fitch framed the experience as evidence that the barrier to building personalised software is falling quickly. Her main lessons were blunt: generated demos can hide mock implementations, moving between local and cloud deployment is now easier than it used to be, and mobile-friendly model design can reduce the need for heavy native dependencies. A GitHub repository linked to the project suggests the app has been published for others to inspect, adapt or extend.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
