Skip to content

Builder project · Built at Pulse Foundry Private Hiring Hack

Harbor-Care-Sot

Combines scattered HR, payroll, license and schedule data into one trusted record, and f

Harbor-Care-Sot project preview
Share on LinkedIn

Tip: click @Pulse AI in the post to tag us.

From idea to reality

What we built

A reusable single source of truth for Harborview Care Group. It takes the four files the business already has (HR roster, payroll, license verification and the printed staff schedule), works out which records belong to the same person, builds one trusted record per employee, and flags everything that is wrong, inconsistent or needs a human to review. Code: https://github.com/Midhili2003/harborview-care-sot Live app: https://harborview-care-sot.streamlit.app/ Working: Recognises each file from its columns, not its name. Each file is scored against every source described in the config, so renamed or reordered columns, Excel files and different date formats still load. The schedule PDF is read page by page: the facility from the title, and shift lengths from the note at the bottom. Cleans every value. "REYES, SOFIA" and "Sofia Reyes", "BYS" and "Harborview Bayside", "Registered Nurse" and "RN" all become one standard form. Nicknames (Marc and Marcus) and accents are handled. Keeps every raw row with its source, so any number can be traced back to the exact file and row. Matches people in order of certainty: HR defines who exists, then license numbers, then name similarity with facility and role as supporting evidence. Uncertain matches are never merged silently; they go to a human. Builds a trusted record where each field comes from the system that should own it (license expiry from the license service, phone from HR, hours paid from payroll), with disagreeing values kept visible. Runs about 25 configurable checks within files, across systems and over time: people paid but not in HR, role or facility conflicts, scheduled vs paid hours, expired or expiring licenses, people working after their license expired, and stale verification. Each issue has a severity, a plain-English explanation and links to its source rows. Remembers human decisions. Confirming a match, marking an issue resolved or accepting it as it is (critical issues require a note) is saved and re-applied on every future ingest.