The Keeper

Most conflicts between people who care about each other are not really about who is right. They are about what never gets said clearly. Keeper is a private AI mediator for two people in a misunderstanding. Instead of forcing both people into the same conversation immediately, Keeper talks to each person separately through a face-to-face video conversation. Each person can explain what happened, how it felt, and what they would need in order to move forward. Keeper does not take sides. It does not decide who is right or wrong. Its job is to listen carefully, understand both perspectives, and find the part where both people already want the same thing. After both private conversations are complete, Keeper compares the two transcripts and identifies the overlap between what each person said they need. Often, people in conflict are closer than they think, but they express the same need in different words. Keeper finds that shared ground and reflects it back clearly. The experience is powered by Tavus Conversational Video Interface, which lets Keeper feel like a real face-to-face conversation instead of a text chatbot. Keeper can listen in real time, respond naturally, and pace the conversation around pauses and emotional tone. The full app is built in Lovable, including the private conversation flow, video integration, transcript processing, and final resolution summary. We built Keeper because the hardest part of conflict is often the silence afterward, when both people feel misunderstood. Keeper does not solve the relationship for them. It simply helps them restart the conversation from the one place they already agree.

Keeper runs two separate, private video conversations: one with each person in the conflict. Each person speaks with Keeper face-to-face, like they would with a neutral mediator. Keeper guides the conversation around three simple questions: What happened? What did it feel like underneath? What would genuinely help resolve it? Keeper follows up naturally, gives people time to think, and avoids interrupting them. It is designed to feel calm, neutral, and emotionally safe. If one person finishes before the other, they see a waiting screen and can return to add more thoughts while the other person is still completing their conversation. Once both conversations are finished, Keeper generates a private summary. The summary shows: The reasoning behind each person’s side What the other person said they are hoping for The specific point where both people’s needs overlap The key output is not advice, blame, or a verdict. It is a clear statement of common ground that both people already expressed, but may not have been able to hear during the conflict.

Keeper uses Tavus Conversational Video Interface for the live video conversation layer. We created a custom Tavus Persona called Keeper with a detailed mediator prompt. The prompt defines Keeper’s role, including neutrality, de-escalation, and the three-part conversation structure. We also tuned Tavus’s conversation behavior for this use case: Raven-1 helps interpret real-time emotional tone. Sparrow-1 manages natural turn-taking. We adjusted the conversation pacing so Keeper waits through pauses and does not rush people while they are thinking. The application itself is built in Lovable as a full-stack TypeScript app using TanStack Start. On the backend, Lovable server functions call the Tavus API to: Create a live video conversation End the conversation when the user is done Retrieve the transcript from Tavus after the call ends Pass both transcripts into the resolution pipeline The final summary is generated with a Lovable AI call using Gemini 2.5 Flash. It reads both transcripts together, identifies what is appropriate to share, avoids exposing private venting, and writes the common-ground summary in one consistent pass. API keys are stored securely as Lovable Cloud secrets and are never exposed in the browser.

Right now, each Keeper session starts fresh. The next major step would be persistent memory. With a real database, Keeper could remember a pair’s conflict history and notice patterns over time. For example, instead of treating every disagreement as isolated, Keeper could recognize that plans, tone, money, chores, or communication delays are recurring sources of tension. We would also explore making Keeper more active during the live conversation itself. Today, the shared understanding appears after both people finish and read the summary. In a future version, Keeper could help surface the common ground live, while the person is still speaking face-to-face with the mediator. Longer term, Keeper could expand beyond couples or friends. The same private-mediation pattern could help with workplace disagreements, family disputes, roommate issues, co-parenting, or any situation where two people each have a side that the other has not fully heard.