For QA reviewers at 503B outsourcing facilities and sterile CDMOs
Every rule-checkable issue in a batch record, found before the investigator finds it.
Sterixia reads an executed sterile fill-finish batch record with its deviation and investigation attachments, and checks it against 60 rules mapped to 21 CFR 211. Your reviewer gets a list of flags, each tied to the cite and the page it is on. It does not make release decisions. Your Quality Unit does.
Become a design partnerThe problem
A sterile batch record runs to hundreds of pages of signatures, timestamps, limits, reconciliations and attachments. One reviewer reads all of it, line by line, often late in the day and before release.
Procedures not followed (211.22(d)) and incomplete investigations (211.192) were the two most cited regulations on FDA's FY2025 drug 483s. Many of those observations are visible in the record itself: a step signed by the same person who verified it, an investigation number with no investigation behind it, a lot number that changes on page 11.
About one check in five needs reading across pages: the reconciliation that disagrees with the step it summarizes, or the deviation on page 12 with nothing attached. These are the issues a tired reviewer misses and an investigator with a fresh copy does not.
How it works
Upload
Your reviewer uploads the scanned or exported PDF, with its deviation and investigation attachments. We check that no page is missing and record the file's fingerprint.
A private reader
An open-weight model reads every page. It runs inside our own AWS account and has no network access, so it cannot send anything out. Every value it reads keeps its page number.
Your reviewer confirms
Each value is shown beside the page image it came from. Your reviewer confirms or corrects it, and every edit is logged.
60 rules, 21 CFR 211
A deterministic rules engine checks the confirmed record. Each rule maps to exactly one cite. Where a cite needs judgment, the flag says so.
The report
Findings grouped by cite family and ordered critical, major, minor. Each one gives the page number and the rule that fired.
What we promise about your records
- Only your people see them. Records are visible only to the people in your organization you invite. Sign-in requires multi-factor authentication.
- Read by a model that cannot send anything out. The reader runs inside our own AWS account with no network. No outside AI service ever sees a record.
- Nobody at Parapet Labs opens them without your permission. Support access is something your admin grants for one record, for a limited time, and you can see that it happened.
- Not used to train anything unless you sign up. If you do, by default the result improves only your own model.
- Deleted when you delete them. Every access to a record is logged, and record content never goes into our logs.
Become a design partner
We are opening a closed beta to a few sterile manufacturers in December 2026. Design partners shape the rules to their SOPs and get early access. We start with synthetic records. Your own records come later, and only under a signed data agreement.