Source notes
Idio
How this project works, the decisions behind it and its known limits. Ask quotes and cites these passages when it answers.
Notes reviewed 20 September 2026
Explore the projectProduct
Product
A keyboard that learns how you type without trying to become your voice. Idio is an iOS keyboard for recovering the words a person intended to type while keeping the person in control. Its core work is local: touch evidence, language context, bounded correction and reversible edits.
The problem it solves
A keyboard can be fast and still be frustrating when it replaces names, slang or intentional wording. A system that only fixes one word at a time can also miss an earlier typo whose meaning becomes clear later in the sentence. The challenge is to use later context without freely rewriting what the person said.
The experience
The keyboard generates candidate corrections, weighs personal and contextual evidence, and can revisit a recent owned span as context arrives. Teach Hands builds a profile for how the person types. Review features provide another bounded opportunity to repair text; Keep Original, protected terms and Undo remain important product behavior rather than premium decoration.
Availability
Idio is in beta on iOS. It is not yet a public App Store release.
Engineering
Backend and architecture
Idio uses native Swift and UIKit for the keyboard and companion app. The local intent decoder combines candidate, language, personal, motor and contextual evidence in a bounded search. Optional review can use Apple's on-device Foundation Models for constrained sentence review, alongside a narrower automatic cleanup path. Model availability is conditional, and deterministic behavior remains the fallback. The keyboard has no cloud dependency for its core correction path; account and sync foundations exist, but they are not released cloud features.
Technology by responsibility
Client: Swift, UIKit, iOS keyboard extension. Intelligence: Local intent decoding, Personal motor evidence, Apple Foundation Models. Platform: App Groups, StoreKit.
Data, trust and failure behavior
Typed text should not enter public telemetry or portfolio screenshots. Local entitlements can project minimal state into the keyboard rather than bringing app-only commerce behavior into a typing extension. Foundation Models can be unavailable, throttled or late; a late result must be rejected when the owned document or revision changes.
Personal touch learning
Teach Hands separates hand mode, speed round and keyboard geometry. Compatible evidence is reused rather than pooled into a misleading single universal accuracy score.
Corrections with provenance
Candidate scores distinguish research, personal and population context. A bounded beam search produces possible readings; an unchanged reading remains meaningful.
Reversible intervention
Protected words, retained originals, owned-document checks and Undo are part of the product contract. A model preference is not permission to bypass those boundaries.
Optional on-device review
It offers explicit sentence review and a narrower automatic cleanup after a pause. Availability, timeouts and fresh-document checks matter; this is not an unrestricted cloud writing assistant.
Decisions
Correcting intent without changing voice
The engine ranks bounded candidate readings and distinguishes personal touch evidence from broader language priors. Protected words and document ownership constrain what can change. Trade-off: A safety boundary can leave a typo unfixed. That is preferable to pretending the system can always know the intended meaning; the user must retain an easy way to correct or undo it.
Learning from trustworthy labels
Teach Hands measures coverage within the relevant hand, speed and geometry channel. The current free-flow calibration regime rejects ambiguous or edited words as strong motor ground truth. Trade-off: Coverage is not accuracy. A larger sample count does not prove better typing, and old calibration evidence cannot silently masquerade as evidence collected under a new labeling regime.
Limits
Known limits
Idio makes no claim to be faster, more accurate or better than Apple's keyboard or SwiftKey; that would need a blinded comparison of correction burden and latency. It is in beta, not a public App Store release. Cloud sync and account foundations exist in the code, but synchronization is not presented as a shipped feature.
Questions Ask answers from here
Does Idio send typing to a cloud model?
The core correction design is local. Optional review uses Apple’s on-device model when available. That does not mean every future account or synchronization feature is already released.
Does Teach Hands measure accuracy?
Its read model reports evidence coverage for a specific hand, speed and layout channel. Coverage and correction accuracy are different measurements.
Why not fix everything automatically?
The intended word is sometimes ambiguous, and a model answer can arrive after the text has changed. Bounded ownership checks and Undo matter as much as a successful correction.