AI-Powered Insurance Estimate SaaS Platform (Next.js + Supabase + AI)

We're building a professional SaaS platform for the property insurance industry, specifically for Public Adjusters (PAs) — licensed professionals who represent property owners in insurance claim disputes (water damage, fire, storm, etc.). The core of the platform: ingest Xactimate PDF estimates (the industry-standard estimating software format used by virtually every insurance carrier and adjuster in the US), use AI to extract every piece of structured data from them, and present it in a clean, professional dashboard. Later phases add a side-by-side reconciliation engine that compares a carrier's estimate against a PA's estimate and surfaces every discrepancy — this is where the real business value lives, but Phase 1 focuses on building the extraction foundation correctly. This is a real multi-tenant SaaS product — not an internal tool. We intend to onboard other PA firms as paying subscribers, and down the line, potentially license access to insurance carriers, contractors, and other professionals in the property claims ecosystem (separate commercial track, not part of Phase 1 — mentioned so you understand the architecture needs to support it later without a rebuild). The Core Technical Challenge (please read before applying) Xactimate PDFs are 30–100+ page structured documents. They look like complex financial reports, but they follow a strict, predictable internal hierarchy: Estimate └── Area (e.g., "Main Level", "Exterior") └── Group (e.g., "Dwelling", "Building Code Upgrades") └── Room (e.g., "Kitchen", "Master Bath") └── Line Items (drywall, paint, flooring, roofing, etc.) Each line item carries a Category Code, Selector Code, Activity Code, description, quantity, unit, unit price, O&P, tax, RCV, depreciation, and ACV. The financial totals roll up through room → category → coverage → estimate level with strict arithmetic relationships. Why this is hard: Standard PDF text extraction returns individual character spans with x/y coordinates — table columns get scrambled on extraction. You need coordinate-based row reconstruction (grouping spans by y-position, sorting by x) before AI can make sense of the data. Then AI handles the semantic interpretation: normalizing codes, classifying line items, and extracting financial values into structured JSON with confidence scoring. This is a solved engineering problem with the right approach — but it requires a team that has actually worked with structured document extraction before, not just "I'll throw the PDF at ChatGPT." What's In Scope for Phase 1 (Fixed Price) User authentication & multi-tenant accounts — registration, login, organization/team accounts, role-based access (owner, admin, adjuster, reviewer, viewer) Subscription billing — Stripe integration with tiered plans, usage metering, customer self-service portal PDF upload & processing pipeline — drag-and-drop upload, virus scanning, coordinate-based text extraction, OCR fallback path for scanned documents AI extraction engine — two-pass extraction (header/metadata pass + line-item pass) using an LLM API of your recommendation (see Tech Stack section — we have a preference but are open to your input) Structured estimate dashboard — multi-tab view showing the estimate broken down by room, by trade category, by coverage, with full financial summary and QA arithmetic validation (does the math actually add up to the printed totals?) Claim management — one claim can hold multiple estimates/versions over time Export functionality — PDF report generation, XLSX export of line items Tech Stack (Our Research — Open to Your Recommendations) Based on our research, here's what we believe is the right stack. However, we are genuinely open to alternative recommendations — if your team has a better-proven approach for any layer (especially the PDF/OCR pipeline or the AI orchestration layer), tell us in your proposal with your reasoning. We care more about long-term maintainability and scalability than about us being "right." Frontend: Next.js 14+ (App Router), TypeScript, Tailwind CSS, shadcn/ui Backend: Next.js API Routes (or your recommended alternative — Fastify, tRPC, etc.) Database: Supabase (PostgreSQL + Auth + Storage + Realtime + Row-Level Security) AI/LLM: We're leaning toward OpenRouter as a model-agnostic gateway (access to Claude, GPT-4o, Gemini, etc. through one API) rather than locking into a single vendor — but open to direct API integrations if you have a strong reason PDF Processing: Python microservice (PyMuPDF + pdfplumber + OCR) Payments: Stripe Deployment: Vercel + Railway/Fly.io for the Python worker Future Phases (Not Part of This Fixed-Price Scope — For Context Only) Phase 2: Two-estimate reconciliation engine (carrier vs. PA line-by-line comparison, AI-powered supplement letter generation) Phase 3: Portfolio analytics, carrier pattern intelligence, photo/material analysis Phase 4: Mobile app (React Native/Expo) We're describing these so you understand the architectural direction — Phase 1 deliverables should not require a rewrite to support these later. Teams that perform well on Phase 1 will be first in line for these phases. Who Should Apply Agencies or senior full-stack pairs with proven experience shipping multi-tenant SaaS products Real experience with LLM API integration for document/data extraction — not just chatbot wrappers Experience with PDF parsing for tabular/financial documents (insurance, accounting, legal, or similar structured-document domains are a strong plus) Comfortable with Supabase or PostgreSQL + Row-Level Security Senior-level TypeScript/React Application Requirements To be considered, please submit: 2–3 relevant projects — links preferred. SaaS products with multi-tenancy, and/or anything involving AI-powered document extraction One paragraph: how would your team approach reconstructing table rows from a PDF where text extraction returns scrambled column order? (We're checking for real experience, not buzzwords) Your fixed-price bid for Phase 1, broken down by week/milestone Any stack recommendations that differ from ours, with brief reasoning Your honest assessment of the riskiest part of this project We will shortlist 3–5 candidates and share a full technical specification document (35-table database schema, complete AI prompts, sample PDFs, design system) before final selection. This is intended as a long-term relationship across multiple phases — Phase 1 is the foundation and the audition.