$ whoami
I turn AI research into trustworthy, production-grade Data & AI strategy.
Düsseldorf, Germany
scrollData & AI Strategist at MultiplAI, helping organisations turn Data & AI investment into trustworthy, production-grade systems, bridging rigorous research and pragmatic delivery.
I hold a PhD (magna cum laude) in computer science from Heinrich Heine University Düsseldorf, where I also worked as a postdoctoral researcher on reinforcement learning post-training and LLM alignment. My research interests span model trustworthiness, uncertainty quantification and, more recently, knowledge graph grounding and agentic systems for trustworthy AI. With 10+ years across academia and industry, my work reaches from publications at NeurIPS, ACL and EMNLP to the open-source tooling that made them possible.
Reinforcement learning and multi-agent coordination for long-horizon decision making in agentic dialogue and tool-using systems.
Reinforcement learning methods for aligning large language models using intrinsic and self-supervised reward signals, reducing reliance on external human preference data.
Reproducible, configurable and distributed training systems enabling rapid experimentation across HPC and cloud environments.
LLM-driven construction and traversal of knowledge graphs (ontology induction from dialogue and natural-language-to-SPARQL/SQL query generation), extending uncertainty-aware methods to entity linking and schema-grounded reasoning for trustworthy, hallucination-resistant enterprise AI.
Bayesian and distributional methods for uncertainty estimation, calibration and robustness, applied to trustworthy and controllable AI systems.
$ tail -n 8 ~/highlights.log
The agentic-skills marketplace carries commits and pull requests, diagnostics, research and Python quality gates. The okf-kb marketplace compiles collected sources into a wiki an agent maintains, with per-source provenance. Eleven plugins between them.
"Leveraging Low-Level Symbolic Competences for Unsupervised Grounding in Hallucination Detection": an LLM builds an SQL database from the reference material and checks each response against it, matching supervised hallucination detectors without any domain-specific training.
"Text-to-SQL Task-oriented Dialogue Ontology Construction": an LLM builds a task-oriented dialogue ontology from scratch using only its SQL capabilities and dialogue-theory concepts in the prompt.
Moving from research to impact, helping organisations take Data & AI beyond the hype and multiply what's possible for their business.
"Less is More: Local Intrinsic Dimensions of Contextual Language Models". The paper uses the geometry of contextual embeddings to study LLM training dynamics and generalisation.
"Post-Training Large Language Models via Reinforcement Learning from Self-Feedback": the model's own confidence becomes the intrinsic reward, with no human feedback required.
"A Confidence-based Acquisition Model for Self-supervised Active Learning and Label Correction": confidence-driven active learning and label validation for sequential multi-output tasks.
Dissertation "Uncertainty Estimation, Management, and Utilisation in Human-Computer Dialogue" defended at Heinrich Heine University Düsseldorf.
$ git log --oneline --career
bd93f9 (HEAD -> now)
Jul 2026 → Present · Germany
ff79c6
Mar 2024 → Jun 2026 · Düsseldorf, Germany
8be9fd
Jul 2019 → Mar 2024 · Düsseldorf, Germany
50fa7b
Jun 2018 → May 2019 · Pretoria, South Africa
Configuration-driven deep learning experiment launcher
Open-source experiment launcher unifying Facebook Hydra, Hugging Face Accelerate and the UV workflow. One config launches anything from a local debug run to multi-GPU distributed training on SLURM clusters or SkyPilot-managed cloud, with no code changes between backends.
Reinforcement Learning from Self-Feedback
Post-training method that uses a language model's own confidence as an intrinsic reward signal, aligning LLMs and improving calibration and reasoning without external human preference labels.
Multi-agent RL for long-horizon agentic systems
Coordinated decision-making between router and expert agents with explicit credit assignment. Write-up in progress.
Workflow, research and code-quality plugins that share one rule: review gates before anything irreversible, and never a bypassed hook.
$ claude plugin marketplace add carelvniekerk/agentic-skills 7 plugins · 14 skills · 4 agents · 3 hooks · 1 MCP server
A knowledge base an agent maintains and you can still trust. Sources stay verbatim, the wiki is regenerated, and every claim traces back to the file it came from.
$ claude plugin marketplace add carelvniekerk/okf-kb $ uv tool install "okf-kb[all] @ git+https://github.com/carelvniekerk/okf-kb" 4 plugins · 12 skills · 1 MCP server
selected work — NeurIPS · ACL · EMNLP · TACL
Carel van Niekerk, Renato Vukovic, Benjamin Matthias Ruppik, et al.
Renato Vukovic, Hsien-chin Lin, Carel van Niekerk, Benjamin Ruppik, Michael Heck, Shutong Feng, Nurul Lubis, Milica Gasic
Renato Vukovic, Carel van Niekerk, Michael Heck, Benjamin Ruppik, Hsien-Chin Lin, Shutong Feng, Nurul Lubis, Milica Gasic
Carel van Niekerk, Christian Geishauser, Michael Heck, Shutong Feng, Hsien-chin Lin, Nurul Lubis, Benjamin Ruppik, Renato Vukovic, Milica Gašić
Benjamin Matthias Ruppik, Julius von Rohrscheidt, Carel van Niekerk, Michael Heck, et al.
Michael Heck, Carel van Niekerk, Nurul Lubis, Christian Geishauser, Hsien-Chin Lin, Marco Moresi, Milica Ga\vsi\'c
University of Pretoria · Jan 2013–Nov 2015 · Pretoria, South Africa
University of Pretoria · Jan 2016–Nov 2016 · Pretoria, South Africa
University of Pretoria · Jan 2017–Nov 2018 · Pretoria, South Africa
Statistical learning, data analytics and visualization.
Heinrich Heine University · Jul 2019–Mar 2024 · Düsseldorf, Germany
Magna cum laude. Thesis: Uncertainty Estimation, Management, and Utilisation in Human-Computer Dialogue.