01. Problem Context & Clinical Constraints
Navigating primary care options is often difficult for patients experiencing unfamiliar symptoms. Without medical training, patients struggle to translate physical sensations into clinical terms or identify the appropriate specialist discipline (e.g., recognizing that persistent fatigue and thirst warrant an Endocrinologist).
Clinical Accuracy & Navigation Latency
The objective was to build an intuitive triage prototype that guides patients toward timely medical consultation without exposing users to open-ended LLM medical hallucinations.
Dual-Engine Triage
Diagnostic triangulation
2.2M+ Records
CMS Data Catalog indexing
Custom KB
Deterministic symptom mapping
02. System Architecture & Triage Pipeline
To prioritize clinical reliability over generative creativity, we implemented a dual-engine architecture combining structured medical databases with constrained LLM validation.
Symptom Extraction
Parses free-form symptom descriptions into normalized clinical entity tokens using structured prompt boundaries.
Diagnostic Triangulation
Queries a deterministic disease-symptom graph for the top 2 candidate conditions, while an isolated LLM call independently hypothesizes a 3rd condition to detect blind spots.
Specialist Matching
Maps triaged conditions to clinical taxonomy codes and queries provider catalogs to return geographic specialist recommendations.
03. Architectural Trade-Offs & Risk Mitigation
Structured Knowledge Base vs. Vector RAG
Evaluated vector search embeddings (ChromaDB) but chose a relational disease-symptom mapping system. This eliminated approximate semantic drift and vector retrieval hallucinations, enforcing deterministic boundaries on LLM context.
Storage Optimization & Query Constraints
Engineered database indexing strategies to handle spatial queries efficiently across the CMS provider catalog while running within constrained memory environments.
Constrained LLM Role
Restricted LLM output strictly to third-party validation rather than autonomous diagnosis, guaranteeing that every clinical recommendation is anchored to verified medical taxonomies.
04. Implemented Prototype & Proposed Architecture
Implemented Triage Prototype
Built and demonstrated a working prototype executing symptom extraction, deterministic candidate matching, and localized doctor routing against CMS provider data.
Proposed Architectural Roadmap
Formulated technical design specifications for automated medical ontology ingestion and dual-factor feedback loops to inform potential future production deployments.