MBILL

After a hospital stay a household receives a pile of bills from providers and a pile of explanation-of-benefits statements from the insurer, and the two piles do not line up. Finding the charge that was billed twice, or billed after the insurer already paid, means a spreadsheet and an afternoon.

The documents are photographs and PDFs. The data on them is some of the most sensitive a person has.

• Upload a bill or a statement as a photo or PDF.
• AI extraction: a vision model reads the provider, dates, codes, amounts billed, allowed, paid and owed.
• The user confirms every extracted field before it is stored. The model drafts; the person decides.
• Matching: bills are matched to statements by provider, date and amount, and the app lists what you were charged that the insurer says you do not owe.
• Encryption at rest for every document and extracted field, and a strict browser security policy.

The document never leaves encrypted storage except to be read by the model and shown to its owner. Extraction is a draft that the user confirms; nothing is matched or totalled on unconfirmed data.

MBILL is in private beta with a small number of households. It is on this page because it is the clearest example of two things we do: AI that drafts under human confirmation, and handling of sensitive personal data as a design constraint rather than a policy document.

Shared infrastructure with its own process and its own encryption keys, monitored by Watchdog. Model calls are rate-limited and never include data from more than one user.