InnerLocket
Private AI-powered photo search using local hardware and semantic search.
Local AI / Edge Computing
Platforms: Web / Local edge hardware
The problem
Find photos by meaning while keeping model inference and the search index under local control.
The goal
Find the photo I’m thinking of, using the words I remember.
Engineering contribution
Private ingestion, semantic retrieval, model integration and index integrity.
Import & duplicate checking
Uploads are staged privately. Visible catalogue entries support duplicate checks before publication.
Checked media processing
The ingestion path validates source identity and authority before publishing decoded media and derived outputs.
Local model execution
The qualified CPU fixture produced 512-dimensional CLIP and 384-dimensional DINO vectors. Production construction expects reviewed local model artifacts.
Catalogue & committed index
Catalogue records, vectors and a database-selected index generation bind retrieval to a coherent library state.
Retrieval & ranking
Natural-language query embeddings feed vector search; lexical search covers filenames and OCR. Rank fusion combines the lists.
Authorized results
Ranked IDs are hydrated from the catalogue while preserving order. Hidden, deleted and vaulted records are filtered out.
Delivery stage: Beta Testing