Running in production
The weekly audit, the board and the monthly exclusion screen run unattended. Caregiver reminders and photo-reply intake are built and being switched on in stages, each with a person approving first.
The situation
A small California home health agency tracked licenses, CPR, TB tests, insurance and annual reviews across its EMR, its personnel folders and a trail of email chases. The three sources disagreed, and nobody trusted any one of them in full.
The cost of leaving it alone
Credentials are the first thing a surveyor asks to see, and an expired license is the most expensive thing to discover late. The EMR’s own “upcoming expirations” report left out items that had already expired, which are exactly the ones you most need to see. A person chasing renewals by email leaves no record you can show anyone.
What we built
A weekly job reads the personnel folders and the roster and applies per-discipline requirement rules, each carrying its regulatory citation. It produces a compliance report, a per-caregiver audit sheet, and a separate task list for the things only the agency can do, such as evaluations and competency checks.
A board in Microsoft Teams shows staff down the side and requirements across the top: green is in date, yellow is expiring, red is expired or missing, and grey means a document is on file but no date has been read yet. Caregivers receive weekly reminders for their own items. The same pipeline runs a monthly screen against the federal exclusion lists and keeps a printable survey record.
Python on scheduled jobsMicrosoft 365: SharePoint lists, Teams, Power AutomateA local vision model on encrypted hardwarePublic federal exclusion data
Why AI, or why not
Mostly not. Dates and set membership are arithmetic. The only AI is a local vision model that reads photos of licenses and cards, and what it reads is a proposal: an office person confirms by reply before anything is filed, because one wrongly read expiry date would mark an expired credential as current. Exclusion matching is exact-name, and a hit is a lead for a person, not a conclusion.
The results
How PHI was protected
- Personnel documents stay on encrypted, agency-controlled storage, and the job refuses to run if that volume is not mounted.
- Photos are read by a local model. No cloud model sees them.
- The personnel folder is mounted read-only and filing is create-only, so nothing is overwritten.
- A person confirms every document before it is filed.
- Reports go only to addresses inside the agency’s own mail domain, and the board is visible to office staff, not to care staff.
- The unattended publisher holds the narrowest SharePoint permission available, for one site only.
What we would tell you to skip
Never let a record the system cannot classify fall through to “not applicable.” The same silent downgrade, a nurse treated as office staff so every clinical rule was skipped, happened three times, and only comparing the output with the real folders caught it. Code review did not.
Reusable for
Small and mid-size agencies with folder-based personnel files and a mix of RNs, LVNs, home health aides and therapists, surveyed by CDPH, CMS or ACHC.