I built the claims database AI risk didn't have.

Insurers are being asked to price AI risk with no loss history to price it from. PerilAI is the missing book: real AI-liability matters, filed by the insurance line they hit, each one carrying its sources.

I designed the data model, the verification standard, the research pipeline, and the product. It is the largest thing I have built.

Visit PerilAI Talk to me about it

↑ the front door

Why it exists Thirteen years on the carrier side of management liability, pointed at a line of business nobody has written the book for yet.

The pipeline

How a claim gets in.

Agentic researchHuman-in-the-loop

Six stages, with my judgment at the two places it actually matters: what counts as an AI loss, and what is sourced well enough to publish.

01Sweep

A weekly crawl of court dockets, regulator actions, and legal press turns up candidate matters.

02Screen

Most candidates are not losses. They get rejected with a written reason, and those reasons become the training set.

03Research

Survivors get the full write-up: what happened, what the AI did, who was sued, and under what theory.

04Coverage read

The part only a carrier person can do: which policy responds, and where the new AI exclusions would bite.

05Review

Every record waits in a queue until I publish or reject it. Nothing reaches the database on the model's say-so.

06Publish

Live to the catalog, the dashboard, and the API at the same moment, sources attached.

↑ my judgment at 04 and 05

Not a deck. A working product.↓ what an underwriter opens

Inside the product

The dashboard

The dashboardEvery kind of AI failure, how often it shows up, and how many lines of coverage each one drags in. The shape of the book before you read a single claim.

The line pageDrill into a line and every claim filed under it opens up, each with the coverage read: which policy responds, and why that distinction decides the claim.

The sources auditEvery citation on one page, tagged primary or secondary, with anything pointing at a search page flagged until it is pinned to a document.

← the coverage read, in my wordsFig. 01 · A line of business, with every claim filed under it.

The standard

How PerilAI decides what's real.

Primary sources firstVisible confidenceNo silent edits

Every record says how sure it isVerified when it is primary-sourced and corroborated, confidence when it leans on reporting, low when the sourcing is not publish grade. Low records stay visible for transparency and are never counted in a headline number.

That rule is the productA database an underwriter will rely on has to be honest about its own edges. It is also the reason I can show you a zero in the unverified column.

Built with Python · FastAPI · SQLite · Claude · agentic research pipelines · a taxonomy I wrote myself

Ayla O'NeillYouTube TikTok LinkedIn Email