A voice worth keeping.

Cadence is an AI conversation companion It uses your writing samples and the context you provide to suggest messages in your style. This paper explains the approach, its tradeoffs, and what still needs to be proven before launch.

01. The problem: a draft without a personality

A message does more than convey information. Its length, punctuation, phrasing, and humour tell the other person who’s speaking. A polished reply can still feel wrong when those details disappear.

Cadence starts with people who overthink dating conversations. The goal is to help them express an intention they already have. Success means a suggestion they recognise as their own, can edit, and feel comfortable sending.

The assistance should make room for your personality.

02. What Cadence learns

You provide examples of your own writing. Cadence builds a style profile: a structured description of vocabulary, sentence length, punctuation, capitalisation, emoji use, and recurring phrases. These characteristics guide the wording of future suggestions.

A profile is an approximation. It cannot capture every joke, cultural reference, or change in mood. Short or unrepresentative samples can produce a poor match. Reviewing and changing your samples helps you shape what the system learns.

03. How a suggestion is made

1. Your examples establish a starting point. Writing samples inform your style profile.
2. You supply the situation. The message or context you paste tells Cadence what you’re replying to.
3. You choose the tone. A casual, flirty, empathetic, or direct direction changes how the suggestion approaches the conversation.
4. A language model drafts the message. The backend combines context, tone, style metrics, and selected writing excerpts into a request to an AI provider.
5. You review the result. You can edit or reject it. Cadence does not send the message to the other person.

The current implementation can include up to 2 recent writing excerpts in generation requests. This gives the model examples beyond abstract metrics, but also means original writing can leave Cadence’s systems. A style profile is personal data; it is not a guarantee of anonymity.

04. Why this approach

Cadence uses a profile and examples rather than training a separate model for each person. That keeps the style context inspectable and lets changes to your samples influence subsequent requests without a new training job.

The tradeoff is imperfect fidelity. Models can overemphasise a phrase, miss your intent, or drift toward generic wording. Adding examples may improve the match, but increases the amount of personal text processed by a provider. Cadence must weigh both outcomes.

05. The boundaries of assistance

AI cannot know what another person feels or guarantee a response. Cadence should not promise romantic outcomes, manufacture a personality, or encourage sending messages you don’t mean.

Generated text may be inaccurate or inappropriate. Your judgement remains part of the process. The site’s interactive example uses fixed illustrative replies; it does not analyse your writing or call an AI service.

06. Data and control

The mobile architecture stores app data locally and synchronises account data with the backend. Saved data may remain available offline; generating an AI suggestion requires a connection. Offline storage does not mean all processing happens on your device.

The important boundary is the provider request. Conversation context, profile information, and selected excerpts may be processed externally. Before launch, the product needs clear provider disclosures, retention rules, and verified deletion behaviour. The privacy notice explains the current development scope.

07. What needs to be measured

The product hypothesis needs evidence from people using it in their own conversations. The evaluation should look at:

• Voice recognition: whether people identify the drafts as something they would write.
• Editing effort: how much they change before they feel comfortable sending.
• Intent preservation: whether suggestions keep the meaning they wanted to convey.
• Control and understanding: whether people understand what is shared with an AI provider and can reject suggestions without friction.

This edition reports no measured results. These are evaluation goals, not evidence that Cadence has already met them.

08. The first release

The first release focuses on learning a writing style, suggesting replies and conversation starters, and choosing a tone. Richer inputs and integrations belong on the roadmap and should earn their place through use.

Cadence starts with a specific moment: helping you move from having something to say to finding words you can stand behind.

Development edition · October 2026. This describes the current approach and open questions, not a promise that every planned feature will ship.