nPlan's Experimental Research Department
We are nERD - nPlan’s Experimental Research Department. We operate as an independent multidisciplinary team, with our long term goals aligned with nPlan’s business objectives. We also publish research contributing to broader scientific audiences in AI/Machine Learning, Project Controls, and Construction Management.
Our mission
nERD’s mission is to advance the state of the art in AI for project controls, enabling robust and predictable projects. We rapidly transfer and deploy innovative technologies into nPlan products ensuring that those products have a real-world impact and stay at the forefront of possibility.
Technical Brief
Our research interests cover a wide range of topics in Machine Learning inspired by our datasets and business needs. For example, our most well-known dataset is composed of over 750,000 construction project schedules, with over 2 billion individual activities. These are DAGs - Directed Acyclic Graphs, where each node represents an activity and edges are constraints between activities. Each node has numerical and textual features, while edges have types, weights and directions. In addition, our graphs have a temporal component representing how each project changed over time.
Doing research at nERD means going deep on Large Language Models (LLMs), Graph Neural Networks (GNNs), Forecasting Science, Stochastic Machine Learning, Uncertainty Modelling, and lots more. We are deeply technical, and extremely rigorous. We solve difficult problems at large scale and make sure that what we say is backed by heavy evidence and strong science.
Another area of active research in nERD is studying how humans interact with, trust and take action using forecasts given by AI. This includes explainable forecasting and creating recommendations for risk mitigation, and generative AI that suggests hundreds of alternative execution and delivery options for our clients to choose from.
Research is in the foundations of nPlan
It is currently done in nERD - nPlan’s Experimental Research Department. We operate as an independent multidisciplinary team, with our long term goals aligned with nPlan’s business objectives. We contribute to shared progress through publication in academic venues.
Recommendations for Risk Mitigation
The first results nPlan’s users see about their projects is the forecasted project end date distribution. We have solved this problem by posing it as an activity duration forecasting task combined with large-scale simulations. The next and more interesting step is recommending mitigating actions. Here we are exploring a range of potential solutions including treatment effect analyses, stochastic schedule optimisation, generation of alternatives, and reinforcement learning. Our first solution, called Intervention Recommender, is currently one of the most powerful tools in nPlan’s project risk mitigation product.
Generative AI and Language Modelling for Projects
During our first few years, we created language models that understand the specifics of construction management. We then created ways to represent any activity from a graph within an embedded space, making use of the textual description, numeric features, and graph structure. This enabled us to research our first forecasting models which are now used in production every day.
Deep Learning on Temporal Flow Networks
In recent years Graph Neural Networks have grown from a niche topic in ML to a prominent and growing research area with a wide range of applications. Our largest dataset is a collection of DAGs where nodes represent individual tasks in the project and edges represent dependency constraints. Although this dataset is unique to nPlan and very different from academic and other industrial applications, we have built highly accurate node and edge forecasting models for nPlan’s product. Our current research interests are around expressive autoregressive models for temporal flow networks.
Humans and AI-driven forecasts
An important area of nPlan’s research is studying how humans interact with AI-driven forecasts.
In particular, we aim to give comprehensive answers to the following questions:
- What makes individuals and teams trust or challenge AI driven forecasts?
- Who are the decision makers in the team and when do they act upon or dismiss AI driven recommendations?
- What is their utility function? How does the presentation of forecasts and risks affect decision makers’ trust in AI?
.webp)
Machine Learning Paper Club
Meet the nERDs
Alan Mosca
Alan spent 7 years as a technologist in quantitative finance, implementing live trading strategies, quotefeed systems, trading engines, writing tools for research and automating front-office tasks. Alan has extensive experience in algorithm design and software engineering, and holds a BEng in ComputerEngineering, MSc in Computer Science, and is nearing a PhD in deep learning
Sophie Gibbs
Sophie holds an MSc in Data Science from City, University of London, where she few-shot generative visual models for tumour detection in 3D images. Since graduating, she has specialised in the rapid prototyping, production design, and deployment of high-accuracy ML solutions for unstructured data across several fast-paced startups.
Inneke Mayachita
Inneke holds an MSc in Petroleum Engineering from Imperial College London and a BSc in Electrical Engineering. She previously worked as a machine learning engineer in oil and gas/facility maintenance industries. At nPlan, she builds the machine learning pipelines and trains the machine learning models that we use to predict risk.
