tl;dr: A developer trained a 125M-parameter decoder-only transformer to autocomplete piano performances in real time (~108 notes/sec on iPhone 15), shipping it as an iOS app called RollTab. Key wins came from a compact note-based MIDI tokenization (pitch, delta, duration, velocity as a single token with summed embeddings), aggressive dataset cleaning over scaling, scheduled sampling between note fields, and DPO post-training using pairwise Gemini preference judgments—which lifted preference rates from ~25% to ~69% over the base model. The model was exported to Core ML with INT8 quantization and uses context rebuilding for long sessions.