Builder project · Built at Pulse Foundry Private Hiring Hack
Pulse SSOT
A single source of truth that reconciles messy HR, payroll, licensing, and scheduling data into one
- ✧ Third Place
Tip: click @Pulse AI in the post to tag us.
From idea to reality
What we built
## The problem Harborview Care Group runs two nursing facilities with 10 employees, but their data lives in four systems that don't agree: HR, payroll, licensing, and scheduling. The same person is written differently everywhere — "REYES, SOFIA" in payroll, "Sofia Reyes" in HR — which makes three things slow and risky: answering referral capacity questions fast, trusting quarterly state staffing reports, and catching an expired license before someone works a shift on it. ## What we built A three-layer pipeline that turns raw, messy exports into one unified, auditable record per employee: 1. **Ingestion & Normalization** — auto-detects CSV columns via fuzzy header matching (no hardcoded schema), extracts tables from PDF schedules with pdfplumber, and cleans names/dates/facility codes into a canonical form. 2. **Entity Resolution** — scores every pair of records across systems using weighted signals (license number match 40%, fuzzy name match 35%, facility + role agreement 25%) and clusters matches above a 0.55 threshold with union-find. Every match stores *why* it happened in plain language — "same license number," "identical name after cleanup" — so nothing is a black box. 3. **Domain Intelligence** — a pluggable layer that turns the unified records into: a weekly capacity calendar that flags shifts worked on an invalid license and suggests who could cover it; a compliance report (real Excel workbook, live formulas, full audit trail) that only compares pay periods the schedule actually covers; and a credential tracker that cross-checks HR's expiry dates against the licensing board's. Disagreement between systems is never hidden — it's surfaced as a conflict, routed to whichever team (HR, Payroll, or Scheduling) actually owns the fix, with a concrete next step. ## Why it's reusable across industries Layers 1 and 2 never reference anything healthcare-specific — no mention of nurses, licenses, or shifts. Only Layer 3 is healthcare-aware, built behind a `DomainPlugin` interface. Swapping in a construction or logistics client means writing a new plugin and updating two config files (facility/job-code aliases, column-detection rules) — the ingestion and matching code underneath never changes. ## What's trustworthy about it - Every number in the compliance report traces back to the exact raw row it came from — this is what "we're never fully confident in [the state report]" actually needs fixed. - License data is resolved by *authority*, not majority vote: the licensing board wins for license fields, HR wins for everything else, because that's who actually owns that fact in reality. - A weak name match alone can never trigger a merge — the threshold is tuned conservative, because a false merge (attributing one nurse's license to a different nurse) is worse than missing a match. ## What we'd build next A real database layer (SQLAlchemy, included), live connectors to HRIS/payroll APIs instead of file uploads, and write-back so an accepted correction pushes back to the source system instead of only being surfaced.