Heza AI: The Sentinel of Rural Health

Executive Command Framework for Autonomous Clinical Intelligence

1. Problem Statement

The primary challenge addressed by Heza AI is the "Clinical Last Mile" in rural Rwanda. Despite having a robust network of 45,000 Community Health Workers (CHWs), the ratio of trained medical doctors to the population remains approximately $1:10,000$. This gap leads to:

• Delayed Triage: Critical symptoms are often dismissed until they become life-threatening.
• Resource Misallocation: Overcrowding of District Hospitals with cases that could have been managed at the health center level.
• Protocol Drift: Difficulty in ensuring strict adherence to Integrated Management of Childhood Illness (IMNCI) guidelines across remote nodes.

2. Proposed Solution

Heza AI is an Autonomous Clinical Triage Agent that operates as a "Senior Analyst" for health. It doesn't just chat; it executes a Self-Correcting Triage Loop $(\mathcal{L})$ defined by: $$\mathcal{L} = {S, A, D, C}$$ Where:

• $S$: Sensing (Multimodal input from CHW/Citizen)
• $A$: Analysis (Cross-referencing symptoms with local seasonal data)
• $D$: Decision (Triage prioritization: LOW $\rightarrow$ CRITICAL)
• $C$: Correction (Autonomous validation of its own reasoning steps)

3. Execution Plan & Methodology

The project was built over a 5-day "Deep Build" period at the University of Rwanda.

• Phase I: Cognitive Mapping: Defining the medical guardrails and local health context.
• Phase II: Architecture Implementation: Building the "Master Gate" security pattern for patient data and the Gemini-powered reasoning core.
• Phase III: Triage Optimization: Engineering the self-correction routines where the agent re-evaluates its risk assessment if vital signs change.

4. Technical Feasibility & Scalability

Heza AI leverages a Vertex AI / Gemini 1.5 Flash pipeline, making it highly scalable across the 3G/4G infrastructure available in Rwanda.

• Scalability Coefficient ($\sigma$): $$\sigma = \frac{\text{Processing Speed}}{\text{Token Density}} \times \text{Regional Impact}$$ The low-latency nature of Gemini 1.5 Flash ensures that even on low-bandwidth rural connections, the clinical "Sentinel" remains operational.

5. Innovation & Usability

The "Wow Factor" lies in our Regional Surveillance Integration. If a CHW in Bugesera reports a spike in fever, the system automatically elevates the "Malaria Probability Weight" ($W_m$) for all subsequent triage events in that specific geographic node. $$W_m = \alpha \cdot (\text{Regional Spike}) + \beta \cdot (\text{Rainy Season Factor})$$

6. Implementation Scenario

1. The Encounter: A CHW enters a household in Musanze.
2. The Pulse: Vital signs and symptoms are fed into Heza AI (Text/Voice).
3. The Audit: The agent identifies an anomaly (e.g., respiratory rate counts don't align with the child's age).
4. Self-Correction: The agent asks the CHW to re-check the rate before finalizing the referral.
5. The Impact: A structured, IMNCI-compliant referral report is generated instantly.

Built for Impact | CBC Hackathon 2026 | UR-CST Node

Built With

• css
• geminiapi
• html
• python
• react
• typescript