Open source

My projects

Code from my research and side quests — crack detection, UAV inspection, embedded sensors, and machine learning for the built world. Pulled live from GitHub.

Featured work

Things I've built and studied

Gaze-informed inspection AI

My PhD thread: recording where expert inspectors actually look during structural assessments, and turning that attention into supervision for vision models.

eye-trackingsaliencyCNNs

Eye-tracking studies with expert building inspectors across facade and disaster case studies — published in Scientific Reports, Buildings, and the Journal of Building Engineering.
Benchmarked saliency-map algorithms against expert visual priorities; used gaze to guide CNN attention, improving generalization to unseen structures.
Optimized architectures for edge deployment: 20% lower inference latency at 85% detection accuracy.

Sci. Reports paper → Research page →

Simulated gaze & lightweight saliency

The active post-PhD extension: when real expert gaze is scarce, simulate it — LLM-designed inspector agents scan a 3D facade, generating training data for compact saliency models.

simulationViT distillationedge AI

End-to-end framework: real PhD fixation data + a 3D concrete facade + LLM-generated inspector personas (junior → expert tiers) drive a Unity simulation of visual inspection behavior.
600 simulated sessions produce fixation sequences, surface heatmaps, and KDE-derived saliency ground truth.
A multimodal ViT teacher is distilled ~8.9× into a MobileNet-class student — under 100MB, sub-50ms inference — then fine-tuned on real gaze for deployment on inspection drones.

Research page → Collaborate →

UAV bridge inspection

Drone imaging with geo-tagging and deep CNNs that turned a slow manual bridge survey into an instant, repeatable visual assessment.

UAVCNNgeo-tagging

Built the end-to-end acquisition pipeline: flight planning, image capture with geo-tagging, and custom image stitching over high-resolution geospatial data.
Data from 50+ inspection flights; real-time deep CNNs on edge devices improved dynamic risk-detection speed by 25%.
Published in Structural Health Monitoring (2021) — my most-cited work.

SHM paper → Related code →

Demand forecasting in production

A production system for forecasting intermittent, hard-to-predict demand — transformer and gradient-boosting ensembles delivering forecasts with uncertainty bands to decision-makers.

TFTLightGBMSageMaker

Took the system from zero to production rapidly, with strong forecast coverage across multi-month horizons for intermittent demand.
Iterated ensembles of temporal transformers and gradient-boosted models, delivering meaningful accuracy gains over the initial baseline.
Owned the full lifecycle — cloud ETL, experiment tracking, automated batch inference, and BI delivery — alongside a document-AI prototype for privacy-sensitive workflows using vision-language models.

Research page → Ask me about it →