AI-Native Property Claims Platform

Property insurers still run claims across fragmented channels — web forms, call centers, adjuster desktops, SIU tools, and payment systems that rarely share one source of truth. Customers expect conversational filing; staff need dense, trustworthy workflows; fraud and payout decisions cannot be left to opaque automation. The gap isn't "add an LLM" — it's an end-to-end platform where conversation, orchestration, and controls share the same claim model, with human gates on money and denials and measurable AI quality after go-live.

Ship a full claims lifecycle — not a chatbot demo — as one coherent product: a nine-status lifecycle with regional role-based access control and fraud holds that block payout, spanning web, voice, and phone intake; a propose-only staff assistant; a rules-plus-LLM fraud engine; Stripe-based payments; and continuous AI-quality operations across a multi-cloud integration fabric (HMAC webhooks, MCP tool auth, Supabase RLS).

• Nine-status claim lifecycle with regional role-based access control
• Fraud holds must hard-block payout regardless of model confidence
• Multiple entry channels (web wizard, voice/text agents, phone) feeding one claim record
• Multi-cloud integration fabric: HMAC webhooks, MCP tool auth, Stripe, Twilio, ElevenLabs, Supabase RLS
• Operations must cover prompt regressions, red-team safety, and input/output drift — not just uptime

A full-stack platform where customers file by form, voice, or phone into one claim model; staff work a role- and region-aware Claims Desktop with an AI assistant that proposes but never executes; a fraud engine blends rules with an LLM workflow to raise explainable alerts and SIU dispositions; approved claims move through Stripe Connect to finance release; and an /observability surface tracks health, quality, and drift with golden and red-team evals gating promotion.

• Next.js on Vercel for the web application; a dedicated Claims-MCP server on Render for long-lived agent tool connections
• n8n Cloud orchestrating the fraud rules + LLM blend workflow
• ElevenLabs for voice/text intake agents; Langfuse for LLM observability
• Supabase as the system of record — guarded RPCs, row-level security, no service secrets in the browser
• Stripe Connect for approve-for-payment → finance release

• Agents propose, humans approve — the desktop assistant can never execute an action
• A fraud hold blocks payout regardless of model confidence
• Every automation run and AI decision is auditable
• Promote to production only behind eval gates: red-team 100%, fraud live accuracy ≥ 85%
• Match hosting to workload shape — long-lived Render for agent tools, serverless Vercel for the web app, rather than forcing one shape for both

• Foundation beats flash — schema, RPCs, and invariants have to precede agent demos, since vibe coding multiplies whatever foundation you give it
• Half the product is contracts — HMAC field order, MCP auth, webhook URLs, and RLS policies fail more often than "the model is dumb"
• MCP needs the right host — long-lived infrastructure for agent tools, serverless for the web tier
• Trust is a feature — observability, golden/red-team evals, and drift monitoring have to ship with the agents, not arrive as a later slide
• Domain review is the control plane — a plausible UI can still be wrong, and humans have to reject unsafe shortcuts even when the code compiles