AIRIS Health Screening Platform
IoT health screening system integrating ESP32-CAM edge devices with a centralized AI processing unit — containerized with Docker and optimized with MLOps pipelines.
- Role
- AI & IoT Developer
- Timeline
- Mar 2025 – Apr 2025
- Team
- 5 people
- Client
- HackFest Universitas Ciputra 2025
The problem
Health screening in resource-constrained settings needs low-cost, portable hardware — but the AI behind it still has to be reliable. AIRIS set out to bring edge-captured health data into a centralised pipeline where every model is validated before it is used.
Constraints
- Low-cost, portable capture hardware (ESP32-CAM)
- Must run the same way across different deployment environments
- Model updates must be validated and reproducible before reaching production
- A hackathon timeline and a five-person team
Architecture
Step 1, Device
ESP32-CAM capture
Portable edge devices
Step 2, Ingest
Data pipelines
Clean and route edge inputs
Step 3, Compute
Inference service
Containerised with Docker
Step 4, Delivery
MLOps validation
Validated, reproducible model updates
Engineering decisions
Containerise the processing stack
- Context
- The system had to behave identically across different deployment environments.
- Decision
- Packaged the inference service and its dependencies with Docker.
A validation gate for every model
- Context
- In a health context, an unvalidated model update is a real risk.
- Decision
- Built an MLOps layer that validates model updates and keeps them reproducible before they reach production.
- Trade-off
- Slower model iteration, in exchange for trust in whatever is running.
Interface contracts between subsystems
- Context
- Five people worked across hardware, ML, and infrastructure at the same time.
- Decision
- Agreed clear contracts for what the camera, the model, and the pipeline expect from each other.
Outcomes
Competition
Infrastructure
Pipeline
- National Finalist at HackFest Universitas Ciputra 2025
- Technical approach and product concept validated in front of a national judging panel
Reflections
Lessons learned
Coordinating a five-person team across hardware, ML, and infrastructure was its own lesson: clear interface contracts between subsystems mattered as much as the code inside each one. When the camera, the model, and the pipeline each knew exactly what to expect from the others, the whole system came together far more smoothly than any individual piece would suggest.
Proof — verify it yourself
- National Finalist — HackFest Universitas Ciputra 2025
Tech stack
- Python
- ESP32-CAM
- Docker
- Deep Learning
- MLOps
- Data Engineering
Work with me
Facing a similar challenge?
I build systems like this end to end — from the first architecture sketch to production. Tell me about yours.