Private AI mediation for people who need clarity before the conversation gets worse.
Most conflict tools help people state their case. Mediator is built to lower the temperature, protect private context, and help both sides arrive at a version of the truth they can actually work with.
Two private AI conversations run in parallel, so each person can speak freely before anything is surfaced
The product is designed to slow escalation down, not reward speed, heat, or performative certainty
An auditor sits in the loop to fail closed on leaks before a response ever reaches the screen
Shipped live with English, French, and Persian support, including RTL
Most conflict resolution tools are built for the mediator, not the people in the conflict. They assume both parties are ready to talk openly, that power dynamics are equal, and that stating your case clearly is the hard part. In reality, the hard part is feeling safe enough to be honest — especially when the conversation is already heated, when one person has more institutional power, or when cultural context makes directness risky.
We built Mediator around one structural idea: two private AI conversations run in parallel. Each person talks to the AI separately, in their own language, at their own pace. They can be contradictory, emotional, unfair — the private channel is a safe space to say the thing you wouldn't say to the other person's face. Only after both sides have been heard does the system begin surfacing shared ground and potential resolutions. An auditor sits in the loop to fail closed on information leaks — nothing reaches the other person that shouldn't.
The AI has to be genuinely neutral — LLMs tend toward agreeableness that can make one person feel validated at the expense of fairness, requiring adversarial testing against dozens of escalation patterns
RTL support across English, French, and Persian isn't just a CSS direction swap — labels, tone, reading direction, and the emotional register of the AI's responses all shift per language
Privacy is load-bearing, not a feature flag — building a fail-closed leak detection system that doesn't degrade the conversation flow was the hardest engineering problem
The product is designed to slow escalation, which means fighting every UX instinct toward speed — deliberate delays, confirmation steps, and cooling-off periods are features, not bugs
Shipped live with trilingual support. People who would not sit in the same room can work through disputes through parallel private channels. The auditor has caught and blocked information that would have escalated conflicts. The design validates the premise: safety before speed, private honesty before public resolution.