PhD Researcher · WMG, University of Warwick

Building the evidence that makes Safe AI trustworthy.

I develop rigorous, interpretable ways to judge whether intelligent systems can be trusted in the real world — with a focus on synthetic-data fidelity and dependable model behaviour for autonomous systems.

  • Interpretable evaluation methods for simulation-to-real transfer.
  • Safety assurance connected to practical validation workflows.
  • Work spanning autonomous systems, computer vision, and high-stakes AI.
See publications Read the blog Get in touch
3 Selected publications
3 Research domains
1 Open-source framework

01

About

A concise overview of the research direction, strengths, and academic positioning behind the work.

Danial Safaei is a PhD researcher at WMG, University of Warwick, within the Safe Autonomy Research Group, investigating how to make AI systems for autonomous settings more trustworthy, measurable, and deployment-ready.

His work focuses on the parts of modern AI that matter most when systems leave the lab: how systems are tested, how synthetic data should be evaluated before it is trusted, and how safety claims can be argued with real evidence rather than appearance.

That agenda spans smart environments, biomedical AI, and autonomous-system validation, but the common thread is consistent: methods that are technically grounded, explainable, and useful in safety-critical practice.

What defines the work

  • Methodological rigor over visual novelty or hype.
  • Research questions tied to real deployment constraints.
  • Clear bridges between theory, evaluation, and safety evidence.

02

Research journey

A progression from applied machine learning toward formal, high-stakes validation for autonomous intelligence.

  1. 2023

    Personalised smart-home intelligence

    Published DeePLT in the International Journal of Information Technology — a deep-learning approach for trajectory-driven adaptive lighting, establishing an early track record in human-centred intelligent systems.

    Springer arXiv DOI
  2. 2024

    Graph learning in biomedical AI

    Co-authored a GraphSAGE study discovering drug synergy in lung adenocarcinoma (the RAIN protocol), showing how graph-based AI transfers into high-stakes scientific and clinical domains.

    medRxiv DOI
  3. 2025

    Decisive Feature Fidelity (DFF)

    First-authored an interpretable framework judging whether synthetic imagery preserves the features that genuinely drive model decisions — validated on 2,126 real–synthetic pairs from KITTI and Virtual KITTI 2.

    arXiv Code DOI
  4. Now

    Safe AI for autonomous systems

    Ongoing doctoral work continues within Safe AI for autonomous systems, building on the evaluation and fidelity methods above toward defensible safety evidence.

03

Research focus

Connected areas of interest across trustworthy AI and autonomous systems.

Primary

Safe AI for autonomous systems

Frameworks that strengthen how safety, reliability, and assurance are argued when AI operates in high-consequence environments.

Testing

Scenario-based testing

Testing autonomous systems against the conditions that matter most for safety and reliability.

Validation

Synthetic-data fidelity

Measuring whether synthetic data is not merely realistic in appearance, but faithful to the decision-relevant structure of real data.

04

Selected publications

Representative outputs across safe AI, computer vision, and graph-based learning.

2025 arXiv Safe AI First author

Quantifying Fidelity: A Decisive Feature Approach to Comparing Synthetic and Real Imagery

Introduces Decisive Feature Fidelity (DFF), an explainable, system-under-test-specific metric that checks whether synthetic images preserve the decision-relevant features real-world performance depends on — validated across PilotNet-style steering and YOLOP segmentation on KITTI / Virtual KITTI 2.

arXiv DOI Code
2023 Springer Smart Environments

DeePLT: Personalized Lighting Facilitated by Trajectory Prediction of Recognized Residents in the Smart Home

A deep-learning framework for adaptive smart-home lighting informed by resident trajectory prediction. Published in the International Journal of Information Technology (vol. 16, pp. 2987–2999).

Springer arXiv DOI
2024 medRxiv Biomedical AI

A GraphSAGE Discovers Synergistic Combinations of Gefitinib, Paclitaxel, and Icotinib for Lung Adenocarcinoma: the RAIN Protocol

Uses graph neural networks to identify drug synergy in lung adenocarcinoma by targeting associated genes and proteins — demonstrating methodology transfer to clinically significant problems.

medRxiv DOI Scholar

05

Open research code

Code that supports reproducibility, transparency, and practical access to the research.

Featured repository

Decisive Feature Fidelity (DFF)

Reference implementation for the DFF framework, supporting transparent evaluation of fidelity between synthetic and real imagery.

Repository Paper

Website source

danial-safaei.github.io

This site is intentionally lightweight: a dependency-free static portfolio in plain HTML, CSS, and JavaScript for fast loading and easy maintenance.

Repository

Full catalogue

All public repositories

Browse the wider body of public work, experiments, and supporting research artefacts on GitHub.

View GitHub

06

Writing

Notes, explainers, and research reflections — published as I go.

The blog

Short, technical pieces on Safe AI, evaluation methodology, and lessons from the research process.

Open the blog

07

Contact

For research collaboration, speaking, or technically serious conversations about Safe AI and autonomous-system validation.

If you are working on trustworthy AI, simulation-to-real transfer, validation methodology, or evidence for autonomous-system safety, I would be glad to connect.

Email danial.safaei@warwick.ac.uk ORCID 0000-0002-4443-8763 LinkedIn danial-safaei Google Scholar Publication profile ResearchGate Research profile GitHub Danial-Safaei