research

Sensing, modelling and reasoning about buildings in use. I work on the instrumented side of the built environment: what a building can be made to report about itself, and what becomes decidable once it does. That runs from sensor infrastructure and edge perception up through forecasting and standards-grounded retrieval.

S.M.A.R.T. Construction Research Group

Visit the lab ↗

NYU Abu Dhabi. Automation and robotics in construction, digital twins for the built environment, AEC cybersecurity, AI, lean construction and BIM, directed by Prof. Borja García de Soto.

Selected projects, published
At the EC³ and CIB W78 conference in Porto, where the paper was published

Predicting indoor particle pollution three days ahead

The scanning robot on the construction site

A robot that scans a construction site on its own

Citations

  1. 2025 · European Conference on Computing in Construction (EC³) · CIB W78, Porto

    Predicting Indoor PM2.5 Levels Using Deep Learning for Enhanced Digital Twin Applications

    A deep-learning model that forecasts fine-particle pollution (PM2.5) indoors three days ahead and runs live inside a digital twin. It beat every model it was benchmarked against, with an average error of 4.9 µg/m³.

    Juan Diego Castaño Molina, Zihao Zheng, Borja García de Soto

    Read the paper ↗

  2. 2025 · Creative Construction Conference (CCC)

    Robotic Systems for Autonomous Data Acquisition in Construction: A Case Study

    A mobile robot that maps an active construction site on its own, then returns to chosen points to take high-resolution laser scans. Tested across 500 m² of a working site with no human at the controls.

    Juan Diego Castaño, Mohamed Benaich, Uday Menon, Samuel A. Prieto, Borja García de Soto

    Read the paper ↗

  3. 2025 · ISPRS Annals · ISPRS Geospatial Week, Dubai

    Digital Twins for Healthier Spaces: A Scalable Framework for Monitoring Indoor Environmental Quality

    A low-cost, self-hosted way to give a building a digital twin that tracks air quality, temperature, noise and light together. It ran for five months with under 3% data loss.

    Zihao Zheng, Juan Diego Castaño Molina, Eyob T. Mengiste, Samuel A. Prieto, Borja García de Soto

    Read the paper ↗
In progress

An air quality assistant that cites its sources

Six-hour-ahead CO₂ forecast against what actually happened

Catching stuffy air six hours before people feel it

Ten campus spaces you can walk through in 3D

Ongoing threads

  • Multi-camera occupancy & activity perception

    YOLOv11m and ByteTrack fused through homography into a single floor-plane view, running at the edge on a Jetson AGX Thor. Set the board up as a shared lab inference endpoint: JetPack 7.0, TensorRT FP16 export, custom Ollama Modelfiles to cap context and work around unified-memory faults.

  • Sensor infrastructure

    TimescaleDB ingestion pipeline and an MQTT broker migration for campus IoT sensors, so downstream work reads from one store instead of five exports.

  • Human factors

    Quasi-experimental study of the IEQ copilot against current practice. NASA-TLX workload effect size d = 0.79; time-on-task d = 0.42.

  • Autonomous reality capture

    A SUMMIT-XL ground robot running LIO-SAM for lidar SLAM and frontier-based exploration, stopping at chosen waypoints to fire a Leica BLK360. Validated across 500 m² of an active construction site; next steps are 3D exploration and lower-latency data handling.

Happy to talk through any of this, or share a draft.

Get in touch