Hello, I'm
Ph.D. Computer Science · Applied LLM Data Scientist · LLMs · RAG · Agents · 75+ citations · EU citizen
I'm an applied LLM and data scientist with a Ph.D. in Computer Science from the University of Groningen (2025) and two years as an applied scientist at Amsterdam UMC's Translational AI Lab. I now work independently on applied GenAI systems. I build LLM-powered and foundation-model systems that solve concrete problems: from domain-adapting an 8B-parameter LLM (Qwen3-8B) with continued pretraining and QLoRA, to retrieval-augmented generation and multi-agent systems, to transformer-based foundation models trained on 1.2M+ clinical records.
I work across the full applied ML stack: pre-training and fine-tuning models on GPU clusters (H100, Slurm), building hybrid-retrieval RAG APIs with reranking and grounded generation, orchestrating LLM agents with Google ADK and LangGraph, and designing the evaluation frameworks and MLOps workflows that take models from prototype to production. Earlier work includes an embedding-based federated learning pipeline deployed at Amsterdam UMC and NHS Blood and Transplant with the University of Cambridge, and GPT-CEP, an LLM pipeline that generates and refines event-processing rules over Kafka/Flink streams.
My PhD and postdoctoral training gave me the depth to design and evaluate these systems rigorously; my focus now is applying that depth to build production-grade LLM and ML products. I've also collaborated with the University of Amsterdam, University of Oslo, TU Ilmenau, and the University of Cambridge.
Almere, Netherlands EU Citizen · No visa required
End-to-end domain adaptation of Qwen3-8B on Amazon Reviews 2023 (131.6M interactions processed): continued pretraining with 4-bit QLoRA on a 1M-document / 108M-token corpus, then LoRA fine-tuning (r = 8/16/32) for rating prediction on H100 GPUs via Slurm. Continued pretraining cut held-out perplexity by 44% (19.5 → 10.9); the evaluation honestly reports where downstream gains did and did not follow.
Conversational product recommender built with Google ADK: a coordinator agent routes between preference-extraction, search, recommendation, and comparison agents. Gemini 2.5 Flash turns natural language into Pydantic-validated preferences; hybrid retrieval (Gemini embeddings + BM25, reciprocal rank fusion) in Qdrant feeds constraint-aware ranking.
Production-oriented RAG service: PDF ingestion and chunking, hybrid dense (BGE) + sparse (BM25) retrieval in Qdrant, cross-encoder reranking, and grounded Gemini answers with source citations, served through FastAPI and deployed with Docker Compose.
LLM system that analyses software repositories and writes architecture, quality, and risk reports. Built twice: as a Google ADK multi-agent system (coordinator + specialist agents) and as a LangGraph pipeline with file-level citations and an LLM review-and-revise step.
Interactive tool that ranks 44 tabular foundation models for a clinical deployment. A clinician questionnaire is translated into weights over 25 sub-metrics across six dimensions (generalizability, privacy, data efficiency, scalability, interpretability, fairness). Runs fully client-side; companion to our comparative framework paper. Try it live →
RAG assistant for technical research papers: page-aware PDF parsing, local MiniLM embeddings in FAISS, and grounded Gemini answers with numbered citations and evidence tables. Declines to answer when the retrieved evidence is insufficient.
Applied LLM system for automated complex event processing rule mining: an LLM reasons over latent data patterns and generates interpretable detection rules, refined through a federated rule-refinement framework over live Kafka/Flink pipelines. Presented at ACM DEBS 2024.
Transformer-based foundation model trained from scratch on 1.2M+ full blood count records. Custom tokenizer, pre-training loop with masked objectives, multi-task fine-tuning heads, and full evaluation harness with W&B tracking.
Continued pretraining and parameter-efficient fine-tuning (QLoRA / LoRA) of open LLMs such as Qwen3-8B, 4-bit quantization, multi-GPU training on H100 clusters, and evaluation that separates language-modelling gains from downstream task gains.
Retrieval-augmented generation with hybrid dense + sparse search, cross-encoder reranking, and grounded, cited answers; multi-agent systems with Google ADK and LangGraph, structured outputs, and tool use.
Transformer foundation models pre-trained from scratch on 1.2M+ clinical records, tabular foundation model benchmarking and selection, representation validation, and interpretable rule extraction for clinical decision support.
Privacy-preserving training deployed across real institutions (Amsterdam UMC, NHS Blood and Transplant) with Flower, embedding-based FL on frozen foundation models, and runtime governance.
End-to-end ML pipelines, FastAPI model services, Docker / Docker Compose, Slurm batch training, experiment tracking with W&B, testing, and reproducible workflows.
Real-time event processing with Apache Kafka and Flink, LLM-assisted rule generation (GPT-CEP), complex event pattern and anomaly detection, and privacy-aware high-throughput pipelines.
Independent GenAI Projects
2026 – Present · Almere, Netherlands
Amsterdam UMC · Translational AI Lab
2024 – 2026 · Amsterdam, Netherlands
University of Amsterdam · IDEAL Lab & Complex Cyber Infrastructure Group
2020 – 2024 · Amsterdam, Netherlands
University of Oslo · Analytical Solutions & Reasoning Group
2020 – 2024 · Oslo, Norway
University of Groningen · Distributed Systems Group
2020 – 2024 · Groningen, Netherlands (defended December 2025)
TU Ilmenau · Distributed Systems & Operating Systems Group
2020 – 2024 · Ilmenau, Germany
University of Groningen, Netherlands
2020 – 2025 · Defended 1 December 2025
Thesis: "Adapting Event Processing in Dynamic IoT Applications: Meeting Evolving Requirements for Quality, Privacy, and Rule Autonomy"
Specialization: Networks & Information Systems for IoT
Research: trade-off between quality monitoring and privacy protection in event processing systems
Foundation in software systems and engineering principles
78 citations · h-index 5 · Google Scholar, October 2026
A Comparative Framework for Evaluating Foundation Models on Tabular Data: A Case Study in Healthcare
arXiv:2609.22154 · 2026 · under review at npj Digital Medicine
From Latent Biomarkers to Clinical Rules: Embedding-Guided Rule Mining and Attribution-Based Translation for Interpretable Tabular Learning
arXiv:2609.22155 · 2026 · under review at Artificial Intelligence in Medicine
A Correspondence Score for Validating Learned Patient Representations in Clinical Tabular Data
arXiv:2608.27489 · 2026 · under review at Journal of Biomedical Informatics
Embedding-Based Federated Learning with Runtime Governance for Iron Deficiency Prediction
FL-ICML Workshop on Federated Learning in the Age of Foundation Models (FLICS) · Accepted 2026 · with University of Cambridge & NHS Blood and Transplant · 1 citation
Towards Federated LLM-Powered CEP Rule Generation and Refinement
18th ACM International Conference on Distributed and Event-based Systems (DEBS) · 2024 · 4 citations
Embedding Space Discovery: Geometry-Based Clinical Data Analysis for Patient Similarity
Under Submission
Patient Classification via Learned Embedding Spaces: Application to Congenital Hypothyroidism Screening
CIBB 2026
Pattern-Level Privacy Protection in Event-Based Systems
SN Computer Science 6 (8), Springer · 2025 · 2 citations
Scheduling Data-driven Workflows in Multi-cloud Environment
Journal of Computer and Knowledge Engineering 1 (1) · 2018
APP-CEP: Adaptive Pattern-Level Privacy Protection in Complex Event Processing Systems
10th International Conference on Information Systems Security and Privacy (ICISSP) · 2024 · Best Student Paper nominee · 5 citations
Driving Towards Efficiency: Adaptive Resource-Aware Clustered Federated Learning in Vehicular Networks
22nd Mediterranean Communication and Computer Networking Conference (MedComNet) · 2024 · 6 citations
AQuA-CEP: Adaptive Quality-Aware Complex Event Processing in the Internet of Things
17th ACM International Conference on Distributed and Event-based Systems (DEBS) · 2023 · 19 citations
Towards Pattern-Level Privacy Protection in Distributed Complex Event Processing
17th ACM DEBS · 2023 · 4 citations
Towards Adaptive Quality-Aware Complex Event Processing in the Internet of Things
18th IEEE International Conference on Mobility, Sensing and Networking (MSN) · 2022 · 9 citations
Scheduling Data-driven Workflows in Multi-cloud Environment
7th IEEE International Conference on Cloud Computing Technology and Science (CloudCom) · 2015 · 28 citations
Released Commerce Foundation Model: end-to-end domain adaptation of Qwen3-8B on Amazon Reviews 2023 with continued pretraining and QLoRA fine-tuning on H100s, cutting held-out perplexity by 44%. Alongside it, a multi-agent recommender built with Google ADK, Gemini, and hybrid retrieval in Qdrant.
View on GitHubThree first-author preprints now on arXiv: a comparative framework for evaluating tabular foundation models in healthcare, a correspondence score for validating learned patient representations, and embedding-guided rule mining that translates latent biomarkers into interpretable clinical rules.
Read on arXivBuilt a series of applied GenAI systems: a production RAG API (hybrid retrieval, cross-encoder reranking, FastAPI, Docker), a paper RAG assistant, and repository intelligence agents in both Google ADK and LangGraph. Also launched the OpTFM Model Advisor, a live tool for choosing tabular foundation models for clinical use.
View on GitHubPresented at the Bloodcount Consortium Workshop at Cambridge's Department of Applied Mathematics & Theoretical Physics. Day 1: presented our federated learning framework for blood count data, built around the real constraints of trusted research environments, using the Flower (flwr) framework with embeddings from the DeepCBC foundation model. Day 2: joined the Foundation Model Workgroup. Ended with a dinner at Trinity College's Great Hall. Great discussions and a strong reminder of why this consortium matters.
View on LinkedInOur paper on embedding-based federated learning with runtime governance for iron deficiency prediction, deployed across Amsterdam UMC and NHS Blood and Transplant, has been accepted at the FL-ICML Workshop (FLICS) 2026, in collaboration with the University of Cambridge.
Read on arXivCurrently open to new applied LLM, GenAI, data science, and foundation model roles, as well as research collaborations. EU citizen, no visa required.