Ian Downard
AI/ML Engineer

Developed an LSTM-based predictive maintenance system for manufacturing equipment, tackling the unique data preparation challenges of time-series sensor data — including labeling strategies, window sizing, and handling imbalanced failure events.

Result: Published methodology on InfoQ covering end-to-end data pipeline design for production ML

Built an image classification model to identify asbestos corrugated sheets, clay roof tiles, and roofing slate from photographs. Deployed to AWS EC2 for automated property defect diagnosis in the UK housing market.

Result: 85–99% accuracy across 3 material categories using 992 training images

Trained an object detection model to identify hand-marked circles on vehicles for automated hail damage counting. Delivered as a Docker image with Jupyter notebook for interactive analysis.

Result: Automated counting of vehicle dents, reducing manual inspection time

Developed a real-time image classification system deployed on Raspberry Pi hardware for identifying surgical instruments (hemostat clamps, needles, scalpels, sponges) with a Gradio web interface.

Result: Near-perfect accuracy with dedicated hardware processing 2 images/minute

Built an object detection model to identify and count specific consumer products on grocery store shelves, demonstrating automated inventory and planogram compliance monitoring.

Result: Real-time product detection across 3 product classes from 212 annotated images

Created a model to classify swimming strokes (freestyle, breaststroke, backstroke, butterfly) from video frames to assist competitive swim coaching.

Result: 85–99% accuracy trained on 7,400+ images across 4 stroke categories

Developed an image classifier to distinguish utility poles containing transformers from those without, reducing manual photo review workload for field inspection teams.

Result: Automated classification of 1,000+ field images into 2 categories