Career Profile
Research Engineer and PhD Candidate at Imperial College London, working at the intersection of NLP, causal evaluation, and AI safety. My dissertation centres on causal-inference methods for NLP, with emphasis on toxicity detection, context effects, and how model judgments compare to human behaviour. In industry, I lead end-to-end NLP and ML projects for internal automation, with a focus on content moderation and localisation.
Experience
Transitioned to Epic Games following the acquisition of Contex.ai in January 2023. Led end-to-end NLP and ML projects for internal automation across multiple business areas, including content moderation and localisation.
- Built and managed datasets for ML applications: data collection, annotation design, quality control, exploratory analysis, and reporting.
- Trained, fine-tuned, and evaluated supervised NLP models — including transformer-based systems and LLM-based approaches — with emphasis on reliable performance in real-world use cases.
- Deployed ML models as production services using Python-based APIs and containerised infrastructure.
- Collaborated with cross-functional teams to define requirements, scope high-impact automation opportunities, document systems, and communicate progress and results.
Contex developed software for content moderation. I led data and research efforts for content moderation systems, with responsibility for textual datasets and contributions to the text-modelling workstream.
- Collected, annotated, analysed, and harmonised textual and visual corpora for model training and evaluation.
- Designed controlled experiments to measure how contextual information affects human judgments of toxicity, and to investigate spurious artefacts in toxicity models using causal methods.
- Contributed to the development of reliable moderation models through data-centric and evaluation-focused research.
Qustodio provides software for parental control. I was part of the product team, responsible for analytics and insights.
- Owned product analytics — defining and tracking product KPIs and SaaS metrics through dashboards and reporting.
- Partnered with product teams to deliver analyses that informed product strategy and A/B test evaluation.
- Applied statistical analysis to generate actionable product insights.
Marie Curie secondment for the ENCASE European Research Programme, conducting research on harmful online content — hate speech, controversy detection, and misinformation.
- Built and analysed a large-scale crowdsourced Twitter dataset for abusive-language research that has been widely used in the field.
- Developed deep learning models for hate-speech and abuse detection.
- Contributed to research on the spread of false or misleading information on social media platforms.
Member of the OSWINDS research group, participating in several research projects:
- Crowdsourcing opinions related to match-fixing and creating a visualisation tool to present project outcomes, for the Erasmus+ European Research Programme Fix the Fixing.
- Studying the impact of external resources on Twitter's information-diffusion process, in collaboration with Zayed University.
Education
- Dissertation focus: causal-inference methods for NLP, with emphasis on toxicity detection, context effects, and the evaluation of model judgments against human behaviour.
- Current work studies whether LLM judges preserve human context effects in toxicity and sarcasm judgments, and develops methods to diagnose and reduce context-insensitive model failures.
Postgraduate programme focused on Data Science, with courses including Applied Machine Learning, Statistical Data Analysis, Web Mining, and (Big) Data Mining. Graduated with an A+ (9.54 / 10).
Publications
- Founta, A.M., Specia, L. "Beyond Majority Labels: Do Multimodal LLM Judges Preserve Human Context Effects on Toxicity and Sarcasm?" Under review at TACL, 2026.
- Madhyastha, P., Founta, A.M., Specia, L. "A study towards contextual understanding of toxicity in online conversations." Natural Language Engineering, 2023.
- Founta, A.M., Specia, L. "A Survey of Online Hate Speech through the Causal Lens." Proceedings of the First Workshop on Causal Inference and NLP, November 2021.
- Founta, A.M., Chatzakou, D., Kourtellis, N., et al. "A Unified Deep Learning Architecture for Abuse Detection." ACM Conference on Web Science (WebSci '19), Boston, USA, June 2019.
- Founta, A.M., Djouvas, C., Chatzakou, D., Leontiadis, I., et al. "Large Scale Crowdsourcing and Characterization of Twitter Abusive Behavior." Proceedings of the 12th International AAAI Conference on Web and Social Media (ICWSM '18), Stanford, California, June 2018. (1,011 citations)
Projects
Awards
Awarded for the winter semesters '15-'16 and '16-'17 of the Postgraduate Programme of the Informatics Department.
Awarded for the presentation of a poster at the womENcourage 2017 scientific event.
Volunteering
SheSharp was a technological community aimed at promoting the engagement of women with computer-oriented technologies, through workshops, hackathons, and other actions.