On-Device Intelligence, and Knowing When to Stay Quiet

The core question behind Mint's classification engine was never "can we sort files?" — it was "when should we sort files, and when should we ask first?" ### The Confidence Problem Most file organizers are binary: either they sort everything automatically, or they ask you about everything. Both fail. Fully automatic systems misclassify files and erode trust. Fully manual systems are just fancy file browsers. We designed a three-tier confidence system: - **Above 92%** — Mint moves the file silently. You will not even notice. - **60% to 92%** — Mint notifies you and waits for a nod. - **Below 60%** — Mint flags it and asks. No assumptions. Getting these thresholds right took more iteration than any other part of the app. We tested with thousands of real files — Downloads folders full of PDFs that could be research papers, invoices, or restaurant menus. The same extension, wildly different intent. ### Four Signals, Layered by Cost Mint classifies each file using four signals, run in order of computational cost: 1. **Extension** — instant, always available, but only tells you format. 2. **Filename patterns** — "Screenshot 2026-" is obviously a screenshot. "Invoice_" is obviously a receipt. Cheap and surprisingly effective. 3. **File metadata** — creation date, size, source app. A 50 MB .mov created at 2 AM is probably not a work document. 4. **OCR content** — the decisive signal. Apple Vision extracts text from images and PDFs on-device. This is what separates a tax form from a takeout menu, even when both are .pdf files. The pipeline runs the cheap signals first. For a folder of 1,000 files, extension and filename patterns classify about 650 instantly. Metadata resolves another 150. OCR handles the remaining 200 — at about 120ms per file on Apple Silicon, the entire folder is done in under a minute. ### Why On-Device Matters We chose Apple's Vision and native language-analysis frameworks deliberately. Not because they were the easiest option, but because they run entirely on your Mac. Your documents never leave your computer. Not during classification, not during OCR, not ever. There is no cloud API, no data pipeline, no server that processes your files. Privacy is not a feature we bolted on — it is a constraint we designed every decision around. ### The Scheduler: Teach It Once The scheduler is built on a simple philosophy: you teach Mint your preferences once, and it remembers forever. Set an organize rule for your Downloads folder. Tell it how you want screenshots handled, where invoices should go, and what to do with disk images after installation. Automatic organization runs only for folders you explicitly configure. The key design choice: automation is owned by each folder. You decide whether it runs automatically, how often it checks, and how Mint organizes that folder. Folders without an automatic rule remain manual. Built on lightweight scheduled checks for only the folders you select. Mint wakes up when a configured folder has work to do, then stays out of the way. ### Learning from Corrections When you correct a classification — telling Mint that a file it called "Scanned Documents" is actually "Medical Records" — it creates a new pattern rule that applies to all future files matching that signature. The system gets smarter with every correction. Not through machine learning or model fine-tuning, but through deterministic pattern matching that you can inspect, edit, and trust.