🤓 Latest Submissions
MARSHALL - Autonomous Insurance Appeal Engine
Every single year in the United States, health insurance companies deny over 850 million medical claims....a staggering 20.4% initial rejection rate. Official government data reveals that over 50% of properly appealed claims are eventually overturned and paid, but less than 1% of patients ever fight back because hiring a healthcare lawyer costs 500 dollars an hour. Insurers count on regular people feeling overwhelmed and giving up. MARSHALL changes the rules by providing an autonomous AI health advocacy engine that turns patient stories into binding legal demand letters. Technical Architecture and 3-Partner Engine: 1. Speechmatics Voice Engine: Real time patient voice dictation, medical entity extraction (CPT and ICD-10 codes), and 98% accurate transcription. 2. Bright Data MCP Scraper: Live 50 state insurance statute scraping in 142ms, prompt-settlement deadline auditing under California Insurance Code Section 2695.7, and fair-market fee benchmarking. 3. AI/ML API Legal Synthesis: Multi-agent legal reasoning, statutory penalty interest calculation, and official Parchment Demand Letter PDF generation for State Insurance Commissioner export. Production and Open Source Standards: Open-source software licensed under Apache 2.0. Built with React 18, TypeScript, Tailwind CSS, Supabase Edge Functions, and Web Speech API.
10 Aug 2026
CORDANE
When a large enterprise reviews a vendor contract, the document is typically passed sequentially between Legal, Finance, Risk, and Operations. Everyone has a different perspective and specific constraints, causing negotiations to stall for weeks. We built Cordane to collapse that timeline into seconds. Cordane is a multi-agent consensus engine built entirely on the Band coordination layer. Rather than using a generic agent swarm, we engineered a strict Mixture of Experts (MoE) architecture. We mapped four specific frontier models to specialized corporate roles: Legal (Claude 3.5 Sonnet) analyzes exposure, Finance (GPT-4o) evaluates margin risk, Risk (DeepSeek Reasoner) ensures global compliance, and Operations (Llama 3 70B) checks SLA feasibility. When a contract is uploaded to the Next.js dashboard, the FastAPI orchestrator initializes a Band Shared Room. All four agents read the document simultaneously. What makes Cordane unique is that they do not work in silos—they negotiate. If Legal flags an uncapped indemnity clause, Finance dynamically reads that task state via Band and immediately recalibrates its margin assessment in real-time. To make this enterprise-ready, we built a strict "Consensus Evaluator" algorithm. Cordane does not force hallucinations just to reach an agreement; if the agents hit a mathematical deadlock on risk tolerance, the system safely halts and escalates the decision to a human executive. Finally, every single API call, constraint check, and agent message in the Shared Room is compiled into an Immutable Audit Log on the UI, giving enterprise governance teams complete visibility into how the AI arrived at its verdict.
19 Jun 2026
