+autocomplete that actually knows you — your vocabulary, your writing style, your recurring phrases, your common recipients, your topics, your opinions. the gap with existing autocomplete (even AI-powered ones like Gmail's smart compose) is that they model generic English rather than your specific voice. the result is suggestions that are grammatically fine but feel wrong — too formal, too casual, using words you don't use, suggesting endings that don't reflect how you'd finish the thought.

+

+the technical approach: train or fine-tune a small language model on your corpus — emails sent, messages, documents, notes — so it learns your voice specifically. the model would autocomplete based on both semantic context (what makes sense to say here) and stylistic context (what sounds like you). a local-first approach is essential: your writing is maximally private data. the inference cost needs to be low enough to run on-device so suggestions appear instantly without a round trip to a server.

+

+this is closely related to [[me-model|me model]] which takes the same personal training data idea further, and to [[axon|axon]] which provides the personal context layer that makes autocomplete smarter. [[embedding-tone-interpolation|embedding tone interpolation]] is a relevant technique — separating tone and meaning in embedding space could let you suggest content that matches the current semantic need while preserving your stylistic fingerprint. [[life-search|life search]] could provide the retrieval mechanism for surfacing relevant prior writing during autocomplete. [[writing-tools|writing tools suite]] is the broader category this falls into.