01. Technical Overview & Problem Context
As an enterprise data systems consultancy, Onix sought to evolve Kingfisher, their synthetic data generation platform. The engineering objective was to transition Kingfisher from relying on static schemas and predefined mock data into a dynamic, prompt-driven agentic generation engine capable of modeling multi-stage relational data on demand.
Healthcare was selected as the pilot domain due to its strict structural specifications and regulatory requirements. While healthcare pathways involve high clinical complexity, patient encounters follow deterministic medical protocols, making it an ideal environment to test multi-agent state machines and schema enforcement.
Engineering Objective
Design a stateful multi-agent system powered by LangGraph to generate structurally sound, FHIR-compliant healthcare records on demand, demonstrating that agentic orchestration can enforce domain-specific constraints without hallucination.
FHIR Compliant
Validated via fhir.resources
20+ Entities
Deterministic patient branching
LangGraph Core
Stateful Agent Orchestration
02. Technical Architecture
Realistic generation required domain-specific knowledge and state tracking. Using LangGraph, I developed an architecture with a semantically enriched knowledge backbone that coordinates specialized agent nodes to simulate sequential clinical outcomes.
Enriched Knowledge Backbone
Curated and normalized a 10,000-record FHIR dataset with relational dependencies and entity mappings across 20+ clinical concepts.
LangGraph State Machine
Coordinates specialized LLM agents and deterministic tools. Graph state maintains shared encounter context to guide conditional branching.
Standardized FHIR Records
Outputs schema-validated FHIR resources modeling end-to-end clinical and insurance encounters on demand.
03. Iterative Implementation Roadmap
Phase 1: Deterministic Baseline
Established baseline validation by coding rule-based state transitions to verify schema constraints across entity types.
Phase 2: Agent-Based Simulation
Transitioned to dynamic simulation using the Mesa agent-based modeling framework to evaluate multi-actor interactions under fixed state rules.
Phase 3a: LangGraph Agentic System
Combined deterministic state graphs with LLM-guided decision nodes, enabling automated intent parsing, dynamic pathway selection, and strict tool validation.
Phase 3b: Insurance Domain Integration
Extended simulation graphs to model insurance workflows, including claims generation, coverage eligibility, and adjudication logic.
04. Evaluation & Technical Impact
Enterprise Technical Validation
Built and delivered technical demonstrators evaluated by engineering and technical leaders across 6 enterprise healthcare and insurance organizations, validating pipeline speed and FHIR schema adherence.
Agentic Architecture Adoption
Validated that cyclic, stateful graphs can replace rigid procedural scripts for structured data generation, establishing the technical foundation to integrate agentic pipelines into Kingfisher's production roadmap.