Experience

Work experience

2024-Present: Postdoctoral Researcher

• Heinrich Heine University, Düsseldorf, Germany
• Dialogue Systems and Machine Learning Group
• Lamarr Fellow Network Ramp Up Project
• Lecturing: Implementing Transformer Models Practical Series
• Agentic Multi-Agent Reinforcement Learning: Led the development of a MARL framework for telephonic dialogue systems, enabling coordinated decision-making between router and expert agents with explicit credit assignment. This approach improved routing accuracy by over 15 percentage points in a production-level dialogue product.

• HydraXcel: Sole developer and maintainer of an open-source, configuration-driven deep learning experiment launcher. Integrated Facebook Hydra, Hugging Face Accelerate, and the UV workflow to support scalable multi-GPU and distributed training. The framework enables seamless, scalable multi-GPU and distributed training for the research team.

• HPC \& Cloud Training Infrastructure: Designed and implemented Hydra launcher plugins enabling transparent execution of experiments on SLURM-managed HPC clusters and via SkyPilot on cloud platforms. This infrastructure enabled high-throughput experimentation and rapid switching between compute backends without code changes.

• Academic Leadership: Supervised multiple Master’s theses focused on MARL and task-oriented dialogue. Designed and instructed the ``Implementing Transformers’’ course, guiding students through building the Attention Is All You Need architecture from first principles in PyTorch, with emphasis on attention mechanics, training dynamics, and debugging. Achieved a 95\% course pass rate.

2019-2024: Doctoral Candidate and Research Scientist

• Heinrich Heine University, Düsseldorf, Germany
• Doctoral Candidate in the Dialogue Systems and Machine Learning Group
• Lecturing: Implementing Transformer Models Practical Series
• Uncertainty-Aware Decision Making: Developed computationally efficient methods for uncertainty quantification in intent classification in collaboration with Yandex Research. Integrated uncertainty features into reinforcement learning policies, improving real-user interaction success by 5 percentage points. Designed an Active Learning strategy that achieved full-dataset performance using only 16\% of expert annotations.
• ConvLab-3 Dialogue Systems Toolkit: Core developer of a large-scale dialogue system toolkit in collaboration with Tsinghua University and Microsoft Research. Architected a unified data format enabling seamless integration of heterogeneous datasets and models, significantly reducing engineering overhead for generalisation and reinforcement learning research. The toolkit has been adopted in 30+ research papers spanning RL- and LLM-based dialogue agents.
• Young Researchers Roundtable on Spoken Dialogue Systems (YRRSDS): Co-organised the 2022 edition collocated with SIGDIAL in Edinburgh. Managed the digital infrastructure and branding for the workshop and contributed to sponsorship acquisition.

2018-2019: AI Consultant

• NGA Risksecure, Pretoria, South Africa
• Consultation on AI solutions
• Implementation of AI solutions
• Projects: Natural Language Processing for Risk Security, Image Classification for plant disease detection

2016-2018: Teaching assistant

• University of Pretoria, Pretoria, South Africa
• Tutorial and Practical Lectures
• Grading of Papers and Assignments
• Modules: Distribution Theory, Stochastic Processes, Actuarial Statistics

Projects

• Uncertainty Estimation in Natural Language Understanding: Designed a framework for estimating and then leveraging uncertainty in dialogue systems to improve user experience and robustness during interactive tasks.
• Dialogue System Evaluation via AWS Mechanical Turk: Designed and developed a custom evaluation interface enabling large-scale human-in-the-loop testing of dialogue agents in AWS.
• Uncertainty Estimation in LLMs: Designed and implemented a custom reinforcement learning framework to improve answer trustworthiness in LLMs.
• Generative AI Feedback Summarization Agent: Designed and implemented a web-based summarization tool that transcribes verbal feedback and generates topic-wise summaries across multiple subjects. Full stack engineered the application using FastAPI and JavaScript with PUG, integrated with OpenAI APIs.
• Entity Sentiment Prediction from Media Reports: Designed, trained, and deployed an AI model to predict entity-specific sentiment analysis from multilingual media sources, enabling real-time sentiment scores for clients in the banking industry.

Professional Activities

• Supervision of BSc and MSc students in the Dialogue Systems and Machine Learning group, 2023 - Present
• Member of the ConvLab 3 toolkit development team, 2020 - Present
• Organizer of the Dialogue Systems and Machine Learning Seminar Series, 2023 - Present
• Reviewer for ACL, EMNLP, NeurIPS, ICLR, AAAI, IJCAI, SIGDIAL, and others, 2020 - Present
• Lecturer and Designer of the Master’s Course “Implementing Transformer Models”, 2023-2024
• Co-organiser of the Young Researchers’ Roundtable on Spoken Dialogue Systems (YRRSDS) 2022
• Supervisor of the Dialogue Reward Estimation Project within the Dialogue Systems and Machine Learning group, 2021-2023
• Lead of the PyDial Python conversion project, 2019
• Co-supervisior of BSc (Hons) Student at the University of the Pretoria, 2017

Service and leadership

• First Tech Challenge Robotics Coach at Woodhill College
• Camera Operator and TV Director at NG Church Moreleta Park
• Camera Operator, TV Director and TV Producer at Hillsong Church Düsseldorf