DEEPSORTENGINE: A LOCAL-FIRST, PRIVACY-PRESERVING INTELLIGENT DESKTOP FILE ORGANIZER USING HYBRID SEMANTIC CLASSIFICATION

Abstract

In the modern digital landscape, desktop users accumulate massive quantities of unstructured files, leading to severe digital clutter and diminished productivity. Traditional file managers lack content awareness, requiring laborious manual sorting or brittle, rule-based configurations. To bridge this gap, this paper presents DeepSortEngine, an intelligent, local-first file organization application that automates file sorting through real-time file system monitoring and a hybrid classification pipeline. The proposed system integrates user-learned patterns, deterministic keyword rules, and deep-learning-based vector embeddings to provide adaptive folder recommendations through a non-intrusive accept/reject user workflow. Crucially, to accommodate deployment on consumer-grade hardware with strict resource constraints, the intelligent engine was migrated from an overhead-heavy PyTorch framework to an inference-optimized ONNX Runtime architecture. This optimization yielded a 99.3% reduction in runtime dependency size (from ~2GB to ~15MB) and an 83% decrease in idle memory footprint (from ~300MB to ~50MB), enabling efficient, CPU-only background operations. Furthermore, the architecture introduces a 7-stage hybrid semantic search engine built directly upon an embedded SQLite vector extension (sqlite-vec), enabling context-rich natural language queries under a local-first, privacy-preserving paradigm.