Re:Using an AI Mediator Because Humans Are Terrible at Conflict
by mksinclair on Monday December 08, 2025 @07:40PM
(#46199838)
Thanks everyone for taking a look — happy to answer anything about the build, the approach, or the decision-making behind it.
A couple notes up front:
Mitigate is currently in pre-launch.
The live demo goes up in a couple weeks — I’ll update this thread the moment it’s ready.
If anyone wants to try it early (or help break it), I opened a waitlist on the site. There are also a few early-access spots for people who want to explore the earliest working version.
Tech details:
Right now the backend is a mix of structured communication modeling, sentiment/tone classification layers, and an LLM-driven mediation pipeline. The hardest parts so far have been:
• keeping the AI strictly neutral (it loves to “fix” the wrong person)
• calibrating tone detection so it doesn’t overreact or underreact
• deciding when to rewrite vs. when to educate
• preventing “therapy drift” — this isn’t meant to replace mental health professionals
Current limitations:
It sometimes errs on the side of too gentle, and nuance in long chains of messages is still tricky. Multi-person conversations (family, teams, etc.) are on the roadmap but not stable yet.
What I’d love feedback on:
• Are the core assumptions here reasonable
• Where could the mediation logic be improved
• Any obvious product traps I’m not seeing
• Ethical/safety concerns around real-time mediation
• Features you wish existed in conflict-resolution tools?
I’m here and will respond as quickly as I can.
Appreciate the time and the critique — genuinely.
A couple notes up front: Mitigate is currently in pre-launch. The live demo goes up in a couple weeks — I’ll update this thread the moment it’s ready.
If anyone wants to try it early (or help break it), I opened a waitlist on the site. There are also a few early-access spots for people who want to explore the earliest working version. Tech details: Right now the backend is a mix of structured communication modeling, sentiment/tone classification layers, and an LLM-driven mediation pipeline. The hardest parts so far have been:
• keeping the AI strictly neutral (it loves to “fix” the wrong person) • calibrating tone detection so it doesn’t overreact or underreact • deciding when to rewrite vs. when to educate • preventing “therapy drift” — this isn’t meant to replace mental health professionals Current limitations: It sometimes errs on the side of too gentle, and nuance in long chains of messages is still tricky. Multi-person conversations (family, teams, etc.) are on the roadmap but not stable yet. What I’d love feedback on: • Are the core assumptions here reasonable • Where could the mediation logic be improved • Any obvious product traps I’m not seeing • Ethical/safety concerns around real-time mediation • Features you wish existed in conflict-resolution tools?
I’m here and will respond as quickly as I can. Appreciate the time and the critique — genuinely.