ML Serving as a Microservice

AI-in-the-Loop Healthcare: Engineering Pregnancy Fit-to-Fly

Pregnancy Fit-to-Fly is not just another full-stack app—it’s a production-ready demonstration of how machine learning can be embedded as a first-class system component, orchestrating user experience in real time.

At its essence, the platform automates the issuance of medical fitness-to-fly certificates for pregnant passengers, but beneath the surface it showcases a robust AI/ML architecture engineered for low-latency inference, modular scalability, and safety-critical decisioning.

---

AI-Driven System Design

Unlike traditional healthcare apps that bolt AI onto the side, this project was built around the principle of AI-in-the-loop workflows. The models are not background advisors—they are the control plane.

The system is composed of three independent but tightly integrated layers:

1. Frontend (clients/): A tab-based user workflow built on modern JavaScript modules.
2. API Backend (server/): Node.js/Express service handling authentication, certificate issuance, and integration with doctors and payments.
3. ML Backend (ml/): A FastAPI-powered inference engine serving lightweight predictive models.

This modular separation ensures that models evolve independently from application logic, enabling rapid iteration and scalable deployment across microservices.

---

Machine Learning Core

The ML service processes structured health data spanning demographics, physiological measures, behavioral scores, and binary symptoms. Feature vectors are dynamically assembled into a numerical pipeline before inference, ensuring data flexibility and consistency.

The models include:

• Heart Rate Predictor (Regression) – fast BPM estimation adjusted for age, activity, and stress.
• Blood Pressure Predictor (Regression) – estimates systolic/diastolic levels from feature interactions.
• Eligibility Classifier (Binary Classification) – outputs probability of “fit-to-fly,” tuned for conservative, safety-first thresholds.

All models are optimized for sub-100ms inference, containerized for cloud portability, and exposed via stateless API endpoints—making them inherently horizontally scalable under load.

---

ML Serving as a Microservice

The ML backend operates as a stateless FastAPI microservice, designed around the following principles:

• JSON schema contracts → Frontend remains model-agnostic.
• Stateless inference → Seamless horizontal scaling via Kubernetes or Swarm.
• API-first extensibility → New models or endpoints can be added without breaking downstream integrations.
• Safety-first gating → Eligibility decisions halt unsafe workflows at runtime.

This design abstracts away complexity for the frontend, while maintaining the flexibility of plugging in new predictive engines (e.g., gestational risk models, wearable-device streams).

---

Production-Grade Stack

The technical stack reflects an engineering-for-scale mindset:

• ML Backend: Python 3.8+, FastAPI, scikit-learn/XGBoost, NumPy, Pandas
• API Backend: Node.js/Express, JWT-based authentication, pluggable data layer (MongoDB/Postgres)
• Frontend: Modular ES modules, esbuild for bundling
• DevOps: Docker-first, .env-based configuration, ready for cloud CI/CD pipelines

Every layer is decoupled but orchestrated, enabling teams to iterate on AI models, backend services, or frontend workflows independently—without regressions.

---

Workflow Orchestration

The application flow illustrates real-time AI integration:

1. Passenger logs in (JWT-authenticated).
2. Health data collected via form-driven interface.
3. Features sent to ML API → model inference controls navigation.

• If “fit-to-fly” → system unlocks doctor selection.
• If not → workflow halts, preventing unsafe certification.
  1. Payment handled via external gateway (Paystack).
  2. Certificate issued as a secure, verifiable PDF.

Here, model predictions aren’t just insights—they are gatekeepers, dynamically shaping user experience.

---

Scalability & Reliability

The architecture is microservice-native:

• Independent deployments → ML and API can scale separately.
• Observability hooks → Request-level logging for monitoring inference performance.
• Security → JWT gating ensures only authenticated requests reach the ML backend.
• Future extensibility → Designed to incorporate explainability, wearables, or federated updates.

In short, this isn’t a prototype—it’s production-ready AI infrastructure applied to healthcare.

---

Future AI-Centric Extensions

The roadmap pushes beyond predictive gating into next-gen applied AI systems engineering:

• LLM Integration: Natural-language medical explanations (“why eligible/ineligible”).
• Edge Deployment: Model exports via ONNX/TensorRT for airline devices or on-prem clinics.
• Explainability Layer: SHAP/LIME to expose feature importance in real time.
• Federated Learning: Decentralized, privacy-preserving model improvements.
• Wearable Integration: Real-time vitals streaming for continuous eligibility tracking.

This trajectory aligns with xAI’s ethos: building real-world AI systems that adapt, explain, and scale.

---

Why This Project Matters

• AI-first architecture – models are core, not add-ons.
• Safety-critical deployment – predictions gate life-impacting workflows.
• Production engineering – decoupled microservices, containerized, and scalable.
• Vision alignment – AI deployed for real human impact in aviation and healthcare.

Pregnancy Fit-to-Fly isn’t just a health app. It’s a blueprint for applied AI systems at scale—a showcase of how machine learning, microservices, and modern deployment pipelines converge into a single mission-critical product.