I Built slothd: A Local-First, Offline-Capable AI Daemon That Patiently Organizes Your Messy Workspace

Description and use

slothd is a local, offline AI daemon that automatically organizes files in a workspace using fast TabPFN metadata triage and deep Gemma 3 document understanding, all without internet access.

Problem

Users spend time manually organizing cluttered workspaces, risking privacy by uploading sensitive files to cloud AI services.

Outcome

slothd organizes 10 test files in under a second, flags 2 clutter anomalies, detects 1 duplicate set, and performs all operations offline with no cloud API calls.

Workflow

• Detect keyboard/mouse inactivity
• Scan target directory
• Extract 7 tabular signals
• Run Prior Labs TabPFN triage
• Flag anomalies
• Run Gemma 3 vision/text model via Ollama
• Generate organization plan
• Show dry‑run preview
• Await user approval
• Execute file moves
• Log actions in SQLite WAL
• Update local vector store
• Provide semantic search via Tiger Data pgvector

Evidence

I built **slothd**, a local-first, offline-capable AI file organizer designed around the philosophy of **"Slow and Steady"**

Source

Uses **Gemma 3** running locally in **Ollama** with strict Pydantic JSON schemas to read PDFs, invoices, and images.

Source

Fast metadata triage (1ms via local 7-signal engine, ~20ms via **Prior Labs TabPFN**) flags anomalies before invoking heavier models.

Source

Every run is a dry-run by default. Every approved move is written to an SQLite WAL log with a unique batch ID.

Source