What we do today, and what we don't have yet.
Fraud Triage handles claim files that name policyholders, claimants and other people. This page lists the controls in place now, in plain terms, and we update it when that changes.
Last updated October 2026.
Claim content is sealed before it's stored.
Encrypted by Lumensic
Uploaded documents, claim narratives, text read from documents, evidence quotes and decision reasons are encrypted by the application with AES-256-GCM before they're stored, on top of our providers' own encryption.
What our key doesn't cover
Details the queue sorts and filters on, such as claim numbers, dates, amounts and statuses, are protected by our providers' encryption only.
In transit
Connections to the app use HTTPS, and the connection to our database is encrypted with TLS.
Only invited people, and only their organization's claims.
Invitation only
No public sign-up. An administrator sets up each organization and invites its people.
Every request verified
Each request to the app carries a session token that is checked before any claim data is read.
Separated by organization
Every claim belongs to one organization, taken from the verified session, and the server checks it on every request. Another organization's claim looks the same as one that doesn't exist.
One shared queue
Everyone in an organization sees its claims. There are no per-person or read-only roles yet.
Who opened what, and when.
What's recorded
Listing, creating, uploading, opening, analyzing and deciding claims, and storing and deleting files, with the verified person, the organization, the time and the result.
What isn't
Narratives, document text, quotes and file contents. The log records which record was reached, never what it said.
Where claim content goes, and when it's deleted.
Two processors
LlamaParse reads text from documents, and Anthropic's Claude API runs the AI review. Real claim data is used only once data-processing terms with both are in place. [CONFIRM: status of both agreements]
Not used for training
Customer claim data is not used to train AI models. [CONFIRM]
Pilot data deleted
Within 30 days of a pilot's end we delete every claim, document, extracted text, flag and decision, including stored files, and send a written confirmation with the counts. The activity log is kept, with file names removed. [CONFIRM: database backups expire inside 30 days]
What we don't have yet.
No SOC 2 report
We have no SOC 2 report or other third-party certification, and we won't imply one.
No roles within an organization
Access is by organization. Per-person, team and read-only roles don't exist yet.
Keys managed by us
Our encryption key is managed and rotated by Lumensic, not by a key-management service.
No log retention period
The activity log is kept until we set a retention period, which is next on our list.
If your review needs any of these, tell us. It shapes what we build next.
Who runs which part.
- VercelHosts the website and the app's interface
- RailwayRuns the application server and database
- Cloudflare R2Stores uploaded documents, encrypted by us
- ClerkSign-in, invitations and organizations
- LlamaParseReads text from PDFs, scans and photos
- AnthropicRuns the AI review of claim documents
Questions for your security review?
Send them to us, along with any questionnaire your team uses. Please don't include claim documents or claimant details.