Patient-Coordinator Matching: Building an AI-Powered Automation System with Supabase Vector Database

The Challenge: Scaling Healthcare Coordination

Healthcare coordination is a complex puzzle. Starling faced a significant operational challenge: 40+ coordinators managing 1000+ patients weekly, each requiring personalized follow-up care. The traditional manual assignment process was inefficient, prone to errors, and couldn't optimize for coordinator expertise or workload balance.

The solution? An intelligent automation system powered by Supabase's vector database capabilities.

System Architecture: Three-Table Foundation

Our solution leveraged Supabase's powerful combination of PostgreSQL and vector search capabilities through a strategic three-table architecture:

Document Metadata Table: Stored essential patient information including risk levels, medical conditions, priority scores, and document types. This table served as the central hub for patient data management.

Document Vectors Table: Housed the vector embeddings generated from patient documents. These high-dimensional representations captured the semantic meaning of patient conditions, symptoms, and medical context.

Coordinator Profiles Table: Maintained coordinator skill sets, specializations, current workloads, and their corresponding skill embeddings. This enabled intelligent matching based on expertise areas.

The foundation of our system relied on Supabase's pgvector extension, which enabled lightning-fast similarity searches across thousands of patient-coordinator combinations.

The Intelligence Layer: RAG-Powered Patient Analysis

Document Processing Pipeline

When patient documents enter our system, they undergo sophisticated analysis. The system extracts medical context, identifies risk factors, and generates comprehensive embeddings that capture not just keywords, but the deeper medical significance of patient conditions.

Our RAG (Retrieval-Augmented Generation) approach ensures that patient analysis goes beyond simple text matching. The system understands medical terminology, recognizes symptom patterns, and assesses risk levels based on comprehensive medical knowledge.

Risk Assessment and Prioritization

Each patient document is automatically analyzed for risk indicators. High-risk patients with complex conditions are flagged for specialist attention, while routine cases are identified for general coordinators. This automated triage ensures critical patients receive appropriate expertise from day one.

Intelligent Coordinator Matching Algorithm

Skill-Based Vector Matching

The core innovation lies in our coordinator matching algorithm that combines vector similarity with business logic. Rather than simple keyword matching, the system understands the relationship between patient needs and coordinator expertise through semantic similarity.

When a patient document is processed, the system searches through coordinator skill embeddings to find the best matches. However, pure similarity isn't enough—the algorithm weighs multiple factors including specialization alignment, current workload, and capacity constraints.

Dynamic Workload Balancing

Our system continuously monitors coordinator workloads and redistributes tasks intelligently. Underutilized coordinators automatically receive appropriate cases, while overloaded specialists are protected from routine assignments that could be handled by general coordinators.

The workload balancing algorithm considers not just current task counts, but also task complexity, coordinator capacity, and historical performance patterns.

Real-Time Assignment Logic

Priority-Based Routing

The system implements sophisticated routing logic that ensures optimal patient-coordinator pairing:

High-Risk Patients are automatically routed to specialist coordinators with relevant expertise. The system applies priority multipliers to ensure these critical cases receive immediate attention from the most qualified professionals.

Medium-Risk Patients are matched with general coordinators who have demonstrated competency in similar cases, with the system considering both skill alignment and current availability.

Low-Risk Patients are distributed among available coordinators with emphasis on workload balancing, ensuring efficient resource utilization while maintaining quality care.

Continuous Optimization

The assignment system operates in real-time, constantly evaluating new patient arrivals against coordinator availability. As coordinators complete tasks, the system automatically assigns new cases based on updated capacity and skill requirements.

Performance Results and Impact

Operational Improvements

The transformation was dramatic. Assignment time dropped from 2-3 hours of manual work to under 30 seconds of automated processing. Matching accuracy reached 94% for appropriate specialist assignments, ensuring patients received care from coordinators with relevant expertise.

Workload distribution became remarkably balanced, with variance across coordinators reduced from ±40% to just ±15%. This balance improved both coordinator satisfaction and patient care quality.

Patient satisfaction scores increased by 23%, reflecting the improved quality of follow-up care when patients are matched with appropriately skilled coordinators.

Technical Performance

The system handles over 100 document uploads per minute while maintaining sub-50ms response times for similarity queries. Supabase's managed infrastructure provides 99.8% uptime, ensuring reliable operation during peak patient intake periods.

Key Technical Insights

Supabase Vector DB Advantages

Supabase's unified platform eliminated the complexity of managing separate vector and relational databases. The ability to combine vector similarity searches with traditional SQL queries in a single system proved invaluable for implementing complex business logic.

Real-time subscriptions enabled instant notifications when new assignments were made, keeping coordinators informed without polling mechanisms. The managed infrastructure removed operational overhead while providing enterprise-grade reliability.

RAG Implementation Benefits

The RAG approach provided context-aware matching that understood patient needs beyond simple keyword matching. As the system processed more patient data, the embeddings continuously improved, creating a learning system that became more accurate over time.

Coordinator expertise could be modeled as vectors, enabling nuanced matching that considered not just stated specializations but demonstrated competencies across various patient types.

Lessons Learned

Database Design Principles

Separating vector and metadata tables proved crucial for optimal performance. While it might seem logical to store everything together, the separation allowed for specialized indexing strategies and query optimization.

Proper indexing strategy became essential as the dataset grew. The system's ability to maintain sub-second response times with thousands of patients and coordinators depended heavily on well-designed vector indices.

Workload tracking emerged as a critical component. Without real-time visibility into coordinator capacity and current assignments, even perfect skill matching would fail to optimize overall system efficiency.

Vector Search Optimization

Embedding quality directly impacted matching accuracy. Investing in high-quality embeddings that captured medical context proved more valuable than optimizing search algorithms alone.

Hybrid scoring that combined similarity measures with business rules consistently outperformed pure vector search approaches. The most effective assignments considered both technical skill alignment and practical factors like workload and availability.

Regular maintenance of vector indices became important as data volume grew. Performance degradation was prevented through scheduled reindexing operations that maintained search speed.

Future Enhancements

The system's foundation enables several exciting enhancements. Predictive analytics could forecast patient no-show probability, allowing coordinators to optimize their schedules proactively.

Multi-modal embeddings incorporating patient demographics, historical data, and outcome patterns could further improve matching accuracy. Real-time learning systems could continuously update coordinator skill embeddings based on patient outcomes and feedback.

Advanced scheduling integration would combine assignment logic with calendar availability, creating a comprehensive coordination platform that manages both task assignment and time allocation.

Conclusion

By combining Supabase's vector database capabilities with intelligent RAG-powered patient analysis, we transformed a manual, error-prone process into an automated, intelligent system. The result: better patient care, optimized coordinator utilization, and significant operational efficiency gains.

The power of vector databases extends far beyond simple similarity search—when combined with domain expertise and business logic, they become the foundation for truly intelligent automation systems. This implementation demonstrates how modern database technologies can solve complex real-world coordination problems while maintaining the flexibility and reliability healthcare organizations require.

The success of this system lies not just in its technical sophistication, but in its practical impact: ensuring that every patient receives follow-up care from a coordinator with the right skills, at the right time, with optimal resource utilization across the entire healthcare team.