NEXSCHED — Healthcare Clinician Matching

Synthesis

A clinician matching system needed to operate under strict clinical and business constraints while remaining fully explainable to administrators and compliant with healthcare regulations.

Architecture

Architected a dual-mode engine combining embedding-based semantic similarity and deterministic rule-based scoring. Multi-tenant architecture with isolated vector indexes ensures zero data leakage across clients.

Terminal State

Deployed with human-in-the-loop workflows and a feedback capture loop. Full roadmap defined for transition to automated learning-based ranking in Phase 2.

The Problem

A US healthcare SaaS platform needed a reliable way to match clinicians to patients under strict business and clinical constraints that generic recommendation systems weren't designed to handle. The matching logic had to be transparent enough for administrators to trust and audit, accurate enough to hold up under clinical scrutiny, and flexible enough to evolve as the platform scaled.

Getting this wrong had real consequences. A mismatched clinician isn't just an operational inefficiency — it's a clinical risk.

What Was Built

A production-grade clinician-patient matching engine with two operating modes — giving administrators full control over the matching methodology depending on their operational and compliance requirements.

The first mode used embedding-based semantic similarity via Sentence Transformers and FAISS — enabling intelligent, context-aware matching across clinician profiles and patient requirements. The second mode used deterministic rule and weight-based scoring with feature hashing — providing full transparency and auditability for compliance-sensitive environments.

The system was built multi-tenant from the ground up. Each tenant operated with ephemeral, isolated vector indexing — ensuring complete data separation across the platform's client base with no risk of cross-tenant data exposure.

Human-in-the-loop workflows kept administrators in control of final matching decisions throughout Phase 1 — building trust in the system's output before any automation was introduced. A feedback capture loop logged every match decision and outcome, laying the data foundation for Phase 2 transition to learning-based ranking models.

The full Phase 2 roadmap was defined and documented — regression-based ranking, automated retraining pipelines — while Phase 1 remained fully rule-governed for reliability and regulatory confidence.

What It Delivered

  • Production-grade matching engine operating under strict clinical and business constraints
  • Dual-mode system giving administrators full control over the matching methodology
  • Multi-tenant architecture with isolated vector indexing per client — zero cross-tenant data exposure
  • Human-in-the-loop workflows ensuring trust, compliance and auditability at every decision point
  • Feedback loop capturing match outcomes as training data for future model evolution
  • Fully defined and documented Phase 2 roadmap for transition to automated learning-based ranking

    • Technology Stack

    PythonFastAPI.NET BlazorSentence TransformersFAISSPostgreSQLDockerMLflowMLOps

    Start Your Project

    You have a problem worth solving. Tell us what it is.

    Describe the problem, not the solution. We will figure out the solution together.

    Let's Talk