Southlake Agentic Synthetic Data Factory

By Qi Sun and Jia An
Advisor: Kaiyu Li

Ivey × Southlake Health Hackathon 2026

Hackathon project focused on AI-driven healthcare workflows, structured data integration, and prototype system development. The team’s participation was recognized through a follow-up $2,500 scholarship opportunity from Ivey MSc.

An agentic synthetic data creation and validation service for healthcare — built for the Ivey × Southlake Health Hackathon 2026.

One-liner: Turn natural language requirements into auditable, FHIR-mapped healthcare datasets — and prove they work through real hospital operation simulations.

Core Features

Feature	Description
Synthetic Data Agent	10-step agentic pipeline: NL → Contract → Strategy → Generate → Audit → Repair → Pack
Data Contract Builder	GPT-4o-mini parses natural language into structured schema, constraints, quality targets
Strategy Selector	Agent chooses generation method based on constraints (Blender / Distribution / Lifecycle)
Cohort Blender	Constraint-weighted resampling with hard constraints (escalation rate) + soft constraints
Dual Audit	Pass 1: Structural (schema, temporal, enum). Pass 2: Plausibility + privacy (k-anonymity, ε, membership)
Repair Loop	Auto-fix temporal inconsistencies, escalation drift, near-duplicates. Before/after metrics
FHIR R4 Mapping	20 fields mapped to HL7 FHIR resources (Patient, Encounter, Condition, CarePlan, etc.)
Simulation Validation	Generated cohorts run through triage/follow-up/transport engines to prove downstream utility
Audit Pack Export	CSV + contract + FHIR dictionary + trace + 3 audit reports + risk statement

Quick Start

pip install -r requirements.txt

# Set API keys
export OPENAI_API_KEY="sk-..."       # Contract generation + Live Doctor
export ANTHROPIC_API_KEY="sk-ant-..." # Triage/FollowUp AI modules

streamlit run app.py

Accounts

Account	Password	Access
admin123	admin123	Synthetic Data Agent (main demo)
admin	admin	Simulation & Validation workspace
user1/2/3	123456	Downstream patient simulation modules

Architecture

Layer 1: CORE — Synthetic Data Agent (the product)
  NL Input → Data Contract → Strategy Selection → Cohort Blender
  → Structural Audit → Plausibility Audit → Privacy Audit
  → Repair Loop → FHIR Mapping → Audit Pack

Layer 2: VALIDATION — Downstream Simulation (proves utility)
  Registration Simulation | Triage Validator | Follow-Up Simulation
  Queue Stress Test | Operations Validation

Layer 3: CONTEXT — Southlake's Real Problems
  $14.8M deficit | 63% labour costs | Aging population
  ED overcrowding | Unsafe discharge | DHN transition

Project Structure

southlake/
├── app.py                              # Login + navigation router
├── core/
│   ├── synth_contract.py               # NL → data contract
│   ├── synth_strategy.py               # Strategy selection with reasoning
│   ├── synth_repair.py                 # Audit repair loop (D0→D1)
│   ├── triage.py                       # Triage scoring (simulation engine)
│   ├── followup.py / routing.py / transport.py / queue_manager.py
├── utils/
│   ├── ai_client.py                    # Claude + GPT-4o-mini clients
│   ├── auth.py / storage.py / styles.py / chat_store.py
├── views/
│   ├── 10_🧬_Synthetic_Data_Agent.py   # MAIN: Agentic pipeline
│   ├── 1-4: Patient simulation modules
│   ├── 5-9: Validation & stress test modules
├── data/
│   ├── patient_records.csv             # 3,000-record seed dataset
│   ├── runs/                           # Generation run history

Demo Video

Project demo videos are available in the Releases section:

• Demo 1
• Demo 2

Safety & Compliance

• ⚠️ Synthetic ≠ Anonymous — model memorization can leak data. Always audit.
• ⚠️ Privacy-Utility Tradeoff — more noise = better privacy, worse fidelity.
• ⚠️ Never for Clinical Decisions — testing, training, simulation only.
• 📋 Regulatory: PHIPA (Ontario) | PIPEDA (Canada) | NIST SP 800-188