Piano Autocomplete

What it does

Piano Autocomplete is an iPhone and iPad app that continues piano performances in real time using an on-device AI model. A musician plays a few notes on a MIDI keyboard connected to their device, and the transformer model predicts what comes next in the piece, generating additional notes automatically. The model processes around 108 notes per second on an iPhone 15 and runs entirely on-device without requiring internet connection.

Who it is for

The app targets MIDI keyboard players and composers who want AI-assisted continuation of their musical ideas. It works as a creative tool for musicians experimenting with melody and performance suggestions, similar to how code autocomplete tools assist programmers. Users need a MIDI keyboard and an iOS device to access the functionality.

Pricing

The app is free.

How it stands out

The implementation uses a 125M-parameter transformer trained specifically on MIDI piano data rather than audio files. The creator emphasizes three technical achievements: finding an effective MIDI tokenization scheme that represents notes, velocities, and timing as separate tokens rather than combined ones; aggressive data cleaning on the training dataset; and Direct Preference Optimization post-training to improve output quality. The model handles the unique challenges of MIDI representation, including enforcing grammatically valid note sequences during generation and managing hanging notes or missing note-off events. Training took approximately one year of iterative experimentation across fourteen different model variants before reaching a satisfactory version.

What a founder should check

First, evaluate whether the on-device constraint is a genuine competitive moat or a limitation. Real-time autocomplete for music exists in other forms, so the defensibility depends on whether musicians actually prefer local processing over cloud alternatives and whether the technical difficulty of running this model size on-device creates meaningful switching costs.

Second, verify the training data moat. The model was trained on public MIDI files, which are widely available. A competitor could potentially achieve similar results with different training approaches, so investigate whether the specific tokenization scheme, data cleaning pipeline, or DPO techniques are genuinely novel enough to prevent replication.

Third, assess the addressable market and pricing pressure. The app is currently free, which removes immediate revenue questions but also suggests the creator is still validating whether musicians will actually use it. Check whether similar music AI tools (code-to-music, music continuation services) are finding sustainable business models or if this remains primarily a creative tool category with low monetization potential.