Personal · on-device LLM · Personal build
itsTypo
Type at full speed; each finished word is checked against the sentence and fixed only when the model is sure.
SwiftmacOSllama.cpp
context
Personal project
The problem
Autocorrect fixes words without understanding the sentence, and stopping to fix typos breaks your flow.
What I built
- A system-wide macOS utility written in Swift.
- At every word boundary, the word you just closed is resolved against your sentence context.
- Runs an on-device language model, so nothing you type leaves the Mac.
- Replaces text only when confidence is high; otherwise it leaves your words alone.
How it works
Watches keystrokes system-wide and detects word boundaries.
Impact
- Private by design: all inference stays on the device.
What I learned
On-device inference lives or dies on latency. Every millisecond between keystrokes counts.