SELF-INITIATED
90-second claim intake
Market opportunity study · policyholder app + adjuster dashboard · designed, built and shipped in 3 days
Project context
The first notice of loss is the most expensive moment in a motor insurer's operation and the one that decides how the customer feels about the policy they bought. It is still handled by phone, with an agent transcribing what someone standing on the shoulder of a highway manages to say.
I took that as a business question and built a working answer: problem framing, design system, two live interfaces and a launch film. No client, no real data. Everything shown runs; nothing shown is claimed as a measured result.
Client
Self-initiated. Market opportunity study
Segment
Motor insurance, first notice of loss (FNOL)
Stage
Working product, publicly deployed. v0 live, v1 architected
Business model
B2B SaaS. Subscription revenue, plans from BRL 200 to BRL 990 per month.
Target audience
•
Policyholder: at the roadside, one hand free, high adrenaline
•
Claims adjuster: at a desk, triaging by severity
Design system
Material Design 3
Role
Product strategy, research, design and implementation
Stack
Claude, Claude Design, Claude Code, Supabase (MCP), GitHub (MCP), Vercel (MCP), Higgsfield (MCP): Seedance, ElevenLabs, Nano Banana, GPT Image, etc.
Period
2026 · 3 days
Challenge
Cut claim intake to under 90 seconds without a phone call, and deliver the data already structured on the other side. The second half carries the value: a form that produces one free-text field only moves the work, because someone still has to read, classify and retype it.
Key actions
•
Design system before screens: Material Design 3 tokens set first, with one rule governing the system; error red reserved exclusively for "someone is hurt". Claim stages use neutral chips, so when red appears it means one thing.
•
Six surfaces in Claude Design: intake, capture, confirmation, live tracking, policyholder home, adjuster dashboard. Delivered as clickable HTML. No Figma at any point.
•
Built with Claude Code and Supabase over MCP: Next.js on Vercel, Postgres and Realtime on Supabase. Schema, RLS policies and Realtime created and verified in the same pass as the code.
•
One write path, by design: every claim goes through a single endpoint. Not the shortest route, and that is the point; the voice layer plugs into the same endpoint and table with no refactor.
•
Honesty in the interface: opening time is measured for real, first tap to server response. The dashboard map carries a "demo data" chip.
•
Launch film with Higgsfield over MCP: 20 seconds inside a fixed 200-credit budget. Cinematic shots in Seedance, generative narration, stills in GPT Image and Nano Banana. App screens are real recordings cut in as inserts, never generated.
Impact
check
Both interfaces live on a public URL, opened on a phone with nothing to install
check
Opening time measured and displayed, never hardcoded
check
v0 complete and presentable on its own, not a draft of v1
Expected impact for a buyer — projections, not results. Self-service intake removes the call from the critical path; structured fields cut first-touch handling time; injuries and non-drivable vehicles arrive flagged from second one; in-app status removes "where is my claim" calls. FNOL is high volume and low unit value, so a modest gain compounds across tens of thousands of claims a year. The buying question is not what the app costs, but how many points of call volume have to leave the contact centre to pay for it.
Launched app
The prototype is now a fully‑launched product. Explore the live app to experience the complete, production‑ready solution.
