Care services · Search and document AI
AI matching and compliance checks for a care marketplace
Matching and compliance checks, built into the marketplace
Worker profiles on the marketplace
Result More than 35,000
The problem
A care marketplace lives on two promises: families find a suitable worker quickly, and every worker on the platform is cleared to work. Filters alone missed good matches. Staff read background checks, screening letters and identity documents by hand, then chased expiry dates. Billing added rules of its own, with rates that change by time of day, at weekends and on public holidays.
What we built
We built the AI service that sits beside the marketplace's core platform. Worker profiles are indexed for vector search, so a family's plain-language request matches on meaning as well as on checkboxes. A vision model reads background checks, screening letters and identity documents and extracts names, outcomes and expiry dates for the onboarding team, and a separate check verifies a worker's business registration. A daily job deactivates workers whose checks have lapsed. For billing, we encoded the time-of-day rules and ran them read-only over past timesheets, so differences were understood before anything changed.
What changed
- A marketplace with more than 35,000 worker profiles
- Families' requests matched to workers by meaning, through vector search over worker profiles
- Background checks, screening letters and identity documents read by a vision model, with key details and expiry dates extracted for the onboarding team
- Workers with lapsed checks deactivated automatically every day
- Time-of-day billing rules tested read-only against past timesheets before any billing change
What production made hard
- Documents arrive as photos, scans and PDFs from many issuers, so extraction reports its confidence and sends unclear ones to a person.
- Deactivating a worker affects their income, so the daily job acts only on clear expiry dates and records every change.
- Billing rules touch money, so they were proven read-only against history before anyone proposed a change.
What a similar project needs
- Sample documents from each issuer you accept, including the messy ones.
- A named person who owns the compliance decisions a model should not make.
- Historical records to test rules against before they touch live billing.
Have a problem in this shape?
Tell us what you are trying to win. We will map your version of this build in one call.