Case studyProfessional
Clinical X-ray AI: from model to production application
My professional work as a Junior Computer Vision Engineer at Unicorn Medics: owning the path from X-ray data and model development to an explainable, containerised application used for clinical review.
- My role
- Junior Computer Vision Engineer: data, models, explainability, application and deployment
- Period
- Jun 2026 – present
- Type
- Professional · Unicorn Medics
The 30-second version
- Problem
- Turn X-ray data into a dependable clinical tool whose predictions clinicians can inspect, and which can be deployed, tested and recovered like any production system.
- My contribution
- Junior Computer Vision Engineer, working across the lifecycle: data preparation, model development, explainability, application engineering and deployment.
- Approach
- Transfer-learning experiments reviewed with clinicians, Grad-CAM explanations, a web application with REST APIs and role-based access, and a containerised, tested deployment.
- Results
- Confidential. Product details, data, model specifics and metrics are not shared publicly.
- Stack
- Python
- TensorFlow / Keras
- OpenCV
- Grad-CAM
- Flask
- PostgreSQL
- Docker
Context
At Unicorn Medics I work remotely as a Junior Computer Vision Engineer on an AI product that analyses X-ray images for clinical review. The role spans the whole lifecycle, so the same person who prepares the data and trains the models also builds and hardens the application those models run in.
Objective
Turn radiograph data into a dependable clinical tool: models that are evaluated carefully, predictions that clinicians can inspect, and software that can be deployed, tested and recovered like any production system.
My responsibilities across the lifecycle
My responsibilities, end to end · select a stage
Stage 1 of 6 · Data
X-ray data preparation
Prepare and analyse radiograph datasets so that models are trained on clean, well-understood data.
- Cleaning and preprocessing
- Exploratory data analysis
- Annotation workflows for detection tasks
Stage 2 of 6 · Model
Model development
Train and compare deep-learning models with transfer learning, and review their outputs with clinicians.
- Comparative experiments across architectures
- Clinician review of model outputs
Stage 3 of 6 · Explain
Explainability
Integrate Grad-CAM so that every prediction comes with a visual explanation a clinician can inspect.
- Heatmaps alongside each prediction
- Supports review rather than replacing it
Stage 4 of 6 · Application
Application engineering
Build the web application around the model, so predictions arrive inside a usable clinical workflow.
- REST APIs
- Database-backed case history
- Role-based access
- Report generation
Stage 5 of 6 · Deploy
Containerised deployment
Package the application and its dependencies so it runs the same way on every machine.
- Docker-based deployment
- Environment-based configuration
Stage 6 of 6 · Harden
Production hardening
Make the system observable, testable and recoverable before it is trusted with real use.
- Health checks
- Automated tests
- Backup and restore procedures
Engineering practices
| Practice | Why it matters in a clinical setting |
|---|---|
| Careful data preparation and EDA | Model quality is limited by data quality; problems are cheaper to find before training. |
| Comparative experiments | Choosing a model is a decision to justify, not a default to accept. |
| Explainability built in | Clinicians need to see why a prediction was made before they can trust it. |
| Role-based access and case history | Different users need different permissions, and every case should be traceable. |
| Containerisation and configuration | Reproducible environments remove “works on my machine” failures. |
| Health checks, tests, backup and restore | A clinical tool must fail visibly, be verifiable, and recover without data loss. |
What this work demonstrates
- Computer vision engineering on real medical images, not only on benchmark datasets.
- Ownership across the stack: the same engineer moves from data and models to APIs, databases and deployment.
- Production thinking: testing, health checks and recovery are part of the job, not an afterthought.
- Working with domain experts: model outputs are reviewed with clinicians rather than judged by metrics alone.
Related public work
The same principles run through my public projects, where the details can be shown in full:
- Diabetic retinopathy screening: full evaluation, confusion matrix and Grad-CAM analysis.
- LungVision: multimodal ensembles with Grad-CAM and LIME, served through a Flask API.
- Network forecasting: tested, containerised Python service with Kubernetes manifests.