AI/ML Engineer for Production Systems
Taking a machine learning problem from raw data to a production system: data pipelines, model selection and training, and deployment infrastructure that holds up under real load, not a notebook that only runs once.
The problem
A model that performs well in a notebook is not the same thing as a system a business can rely on. The gap between an accuracy metric that looks good and a system that runs unattended in production, on new data, without someone babysitting it, is where most ML projects stall. Teams either lack the engineering discipline to close that gap, or the person who built the model isn't the person who can deploy and operate it.
The technical approach
I treat the full lifecycle as one job: data ingestion and validation, model selection matched to the actual problem rather than the trend, and a deployment path decided before training starts. For structured and tabular problems I use Scikit-learn and XGBoost, since gradient-boosted trees often outperform neural networks on that kind of data. For vision and sequence problems I use PyTorch and TensorFlow, with transfer learning as the default starting point rather than training from scratch. Every training run is tracked in MLflow so experiments are reproducible and comparable, and every model is versioned with artefacts stored in AWS S3 so a rollback is always available, not theoretical.
What you get
A deployed FastAPI inference service, containerised with Docker, running on AWS EC2, with a CI/CD pipeline in GitHub Actions that ships model updates without manual SSH sessions. This is the pipeline I built for SmartInsureAI, a vehicle insurance risk-assessment system: automated data validation and feature engineering, a Scikit-learn model served through a real-time FastAPI endpoint, and a full CI/CD path through Docker, Amazon ECR, and EC2, the same infrastructure pattern I use for every model I ship.
Projects that demonstrate this
MLOps · Scikit-learn
Smart Vehicle Insurance Risk Assessment
End-to-end MLOps pipeline for insurance risk scoring: automated data ingestion, validation, feature engineering, Scikit-learn model training, and real-time FastAPI prediction APIs. Full CI/CD with GitHub Actions, Docker, Amazon ECR, and EC2.
Read case study →Computer Vision · Transfer Learning
Wrist Abnormality Detection (YOLOv9)
Fine-tuned YOLOv9 on the GRAZPEDWRI-DX medical imaging dataset for real-time wrist X-ray abnormality detection. Full pipeline from data preprocessing and augmentation through hyperparameter optimisation to a Flask inference API.
View project →Core tools & frameworks
Related services
Need an AI/ML Engineer?
Whether you're building a new model from scratch, productionising an existing research prototype, or need MLOps infrastructure that actually works, let's talk.