TRUST2: Building Explainable AI for Human-Occupied Spaces
Building on the UKRI-funded TRUST2 project developed at University College London, this workshop presents the methods we used to develop AI-driven building management solutions featuring trustworthy and explainable AI.
During the workshop, we will explore the architecture we employed to build end-to-end explanatory Large Language Model (LLM) solutions, guardrail systems, and multi-architecture pipelines that safely manage human-occupied sites.
In particular, we will examine how grounding systems ensure truthful and trustworthy explanations for the decisions made by black-box models that impact humans.
We will also examine newer techniques, such as knowledge graph systems, that can be used to constrain AI behaviours and explain decision processes.
Participants will receive an in-depth overview of the guardrails, multi-model pipelines, and LLM-as-judge systems used in the TRUST2 project. They will apply these methods hands-on to example data on their laptops and see the results in real time.
Attendees will leave with a practical and theoretical understanding of:
- Architectural Frameworks: Methodologies for combining discrete time-series forecasting algorithms with large language models to securely manage physical environments.
- System Guardrails: Techniques for constraining LLM outputs and mitigating emergent behavioural hallucinations to ensure the safe execution of control signals, such as knowledge graphs.
- Knowledge Graphs for Explainability: Application of annotated knowledge graphs to combine black-box LLMs with grounded, explainable reasoning traces.
It is recommended that participants read this summary of the TRUST2 project: https://connected-environments.org/projects/trust2/
We also recommend reading this overview paper on explainable AI: https://dl.acm.org/doi/full/10.1145/3561048
The Speaker(s)
Jack Shiels
Jack Shiels is an AI researcher and founder with extensive experience building and deploying large language models (LLMs) and reinforcement learning systems across research and commercial environments.
He is also the co-founder of shiels.ai, a Malaysian company where he has led the development of AI systems, including an RL-based research impact forecasting agent and an ontology-driven enterprise requirements and agent orchestration platform.
Previously, Jack conducted government-funded AI research at University College London on trustworthy LLM systems for IoT environments, contributing to a model ecosystem that reduced energy wastage by over 80%.
Before moving into AI research, Jack worked as a full-stack fintech software developer in the City of London. Born in Cape Town, South Africa, he has lived most of his life abroad in the UK, Ireland, and Malaysia.
