The latest articles on DEV Community by acuerdio (@acuerdio). https://dev.to/acuerdio https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3869503%2F3f455caa-058f-4d67-9c66-6174086bf5e2.png https://dev.to/acuerdio en acuerdio Thu, 09 Apr 2026 09:50:09 +0000 https://dev.to/acuerdio_es/real-time-mediation-with-websockets-how-we-built-a-legal-dispute-resolution-platform-aaf https://dev.to/acuerdio_es/real-time-mediation-with-websockets-how-we-built-a-legal-dispute-resolution-platform-aaf <p>When two parties are in conflict, communication is everything. At Acuerdio (<a href="https://acuerdio.com" rel="noopener noreferrer">https://acuerdio.com</a>), we built a real-time mediation platform where AI and human mediators guide disputes to resolution — all through WebSockets.</p> <h2> Why Real-time Matters in Mediation </h2> <p>Mediation isn't like a chatbot conversation. It's a structured negotiation where:</p> <ul> <li>Both parties need to see proposals simultaneously</li> <li>A mediator (AI or human) orchestrates the conversation</li> <li>Legal documents need to be drafted and signed in real-time</li> <li>Video sessions happen alongside text-based negotiation</li> </ul> <p>We needed a system that handles all of this with sub-second latency.</p> <h2> Architecture: Django Channels + Remix </h2> <p>We use Django Channels with a Redis channel layer for WebSocket communication. The frontend uses Remix 2 + React 18 with custom hooks for real-time state management.</p> <h2> The Mediation Flow </h2> <ol> <li>Party A opens case</li> <li>AI analyzes conflict</li> <li>Both parties join real-time session</li> <li>AI proposes resolution / Parties negotiate</li> <li>Agreement reached</li> <li>eIDAS signature</li> <li>Legal document generated</li> </ol> <h2> AI as Mediator </h2> <p>The AI mediator participates in the WebSocket session just like a human would — analyzing both parties' positions and suggesting fair resolutions in real-time.</p> <h2> Video Integration with Daily.co </h2> <p>For complex cases, we offer video mediation sessions powered by Daily.co with:</p> <ul> <li>Private rooms (max 4 participants)</li> <li>Recording enabled (legal requirement)</li> <li>1-hour session limit</li> </ul> <h2> Scaling Considerations </h2> <ul> <li> <strong>Redis Cluster</strong> for channel layer (handles 10K+ concurrent connections)</li> <li> <strong>Celery workers</strong> for async LLM calls (don't block the WebSocket)</li> <li> <strong>PostgreSQL with pgvector</strong> for legal document similarity search</li> <li> <strong>S3</strong> for recorded video sessions and signed documents</li> </ul> <h2> Lessons Learned </h2> <ol> <li> <strong>WebSocket reconnection is critical</strong> — Users on mobile networks drop frequently</li> <li> <strong>Message ordering matters</strong> — In legal mediation, the sequence of proposals is legally significant</li> <li> <strong>AI latency management</strong> — Show typing indicators while AI processes</li> <li> <strong>Privacy by design</strong> — All messages are encrypted, mediations are confidential by law</li> </ol> <p><em>Built with Django Channels, Remix, and a lot of legal expertise. Acuerdio (<a href="https://acuerdio.com" rel="noopener noreferrer">https://acuerdio.com</a>) — Resolve disputes without going to court.</em></p> websocket django react realtime acuerdio Thu, 09 Apr 2026 09:48:15 +0000 https://dev.to/acuerdio_es/building-an-ai-mediator-multi-llm-architecture-for-legal-dispute-resolution-a9o https://dev.to/acuerdio_es/building-an-ai-mediator-multi-llm-architecture-for-legal-dispute-resolution-a9o <p>At Acuerdio (<a href="https://acuerdio.com" rel="noopener noreferrer">https://acuerdio.com</a>), we built Spain's first AI-powered online mediation platform. Here's how we designed a multi-LLM architecture that understands legal conflicts and proposes fair solutions.</p> <h2> The Problem </h2> <p>Spain's new law (LO 1/2025) requires parties to attempt alternative dispute resolution before going to court. That's millions of potential cases — from neighbor disputes to family conflicts — that need a scalable, AI-assisted mediation process.</p> <p>We needed an AI system that could:</p> <ul> <li>Understand legal context in Spanish</li> <li>Propose balanced solutions to both parties</li> <li>Generate legally valid documents</li> <li>Handle sensitive emotional situations with empathy</li> </ul> <h2> Architecture: Multi-Provider LLM Strategy </h2> <p>We don't rely on a single LLM. Our architecture uses multiple providers with intelligent routing.</p> <h3> Why Multi-LLM? </h3> <ol> <li> <strong>Resilience</strong> — If one provider goes down, we failover automatically</li> <li> <strong>Cost optimization</strong> — Use cheaper models for simple tasks</li> <li> <strong>Best-of-breed</strong> — Each model excels at different aspects of mediation</li> <li> <strong>Compliance</strong> — We can route sensitive data to EU-hosted models</li> </ol> <h2> The Mediation Pipeline </h2> <ol> <li>Case Analysis → Understand the conflict type and legal framework</li> <li>Party Profiling → Assess each party's position and interests</li> <li>Proposal Generation → Create balanced resolution proposals</li> <li>Agreement Drafting → Generate legally binding documents</li> </ol> <h3> Legal Context Understanding </h3> <p>Spanish legal mediation has specific requirements. We built a RAG system with:</p> <ul> <li>Spanish Civil Code</li> <li>LO 1/2025 (mandatory ADR law)</li> <li>Regional mediation regulations</li> <li>CGPJ case law database</li> </ul> <h3> AI-Assisted Proposals </h3> <p>The AI generates resolution proposals that consider:</p> <ul> <li>Legal precedent for similar cases</li> <li>Both parties' stated interests</li> <li>Fair distribution principles</li> <li>Enforceability of the agreement</li> </ul> <h2> Tech Stack </h2> <ul> <li> <strong>Backend:</strong> Django 5 + DRF + Channels (WebSocket for real-time mediation)</li> <li> <strong>Database:</strong> PostgreSQL 16 with pgvector for embeddings</li> <li> <strong>Task Queue:</strong> Celery + Redis for async LLM calls</li> <li> <strong>Frontend:</strong> Remix 2 + React 18</li> <li> <strong>Video:</strong> Daily.co for live mediation sessions</li> <li> <strong>Signatures:</strong> eIDAS-compliant electronic signatures</li> </ul> <h2> Results </h2> <ul> <li>AI resolves ~70% of simple disputes without human mediator involvement</li> <li>Average case resolution: under 72 hours (vs 14.3 months for courts)</li> <li>Cost: from 9 EUR (vs 3,000-10,000 EUR for litigation)</li> </ul> <h2> What's Next </h2> <ul> <li>Voice-based mediation using real-time speech-to-text</li> <li>Predictive analytics for case outcomes</li> <li>Multi-language support (Catalan, Basque, Galician)</li> </ul> <p><em>Acuerdio is Spain's first AI-powered mediation platform. Try it at acuerdio.com</em></p> ai legaltech python architecture