Building multi-agent AI systems, low-resource NLP pipelines, and voice AI:
from supercomputer-scale data harmonization to interactive Socratic tutoring platforms.
💬 redpill47.github.io — my portfolio, styled as a chat with Claude. Ask it anything about me.
Currently at DCS, University of Luxembourg: building LLM annotation pipelines and RAG evaluation benchmarks for Luxembourgish, a low-resource language. Pipelines orchestrate OpenAI, Anthropic, and DeepSeek APIs with self-hosted Langfuse observability on Docker. Freshly graduated from the BiCS bachelor (2023–2026), starting an MSc in Information and Computer Sciences at the University of Luxembourg this autumn.
At LISER (Mar–Jun 2026): Architected the MIGLINK University Harmonization pipeline: mapped 594M LinkedIn education records to ROR-standardized identities across 63M+ raw institution strings on the MeluXina supercomputer (SLURM). Cascading matching (exact → suffix-normalized → LLM) achieved 100% precision / 94.8% recall using Qwen3-32B via vLLM.
At AI Robolab, Uni Luxembourg (Sep 2024–Mar 2026): Led development of a Luxembourgish AI tutoring app, designed open-source LLM benchmarking pipelines, co-authored 5 peer-reviewed papers, and represented the lab at CALM '25 demoing Furhat and QTrobot.
| Venue | Title | Link |
|---|---|---|
| ACM CUI '26 | How Authoritative Can Synthetic Voices Sound? Prosodic Manipulation and Perceptual Evaluation Across Two TTS Platforms | DOI |
| AAMAS '25 | Personalized Language Learning: A Multi-Agent System Leveraging LLMs for Teaching Luxembourgish | |
| MobiSPC '25 | Artificial Intelligence Agents for Personalized Adaptive Learning | DOI |
| MDPI Information '25 | BPMN-Based Design of Multi-Agent Systems: Personalized Language Learning Workflow Automation | DOI |
| EUSPN '25 | Can Visual Attention Serve as an Explanation? | DOI |
| CALM '25 | Augmenting an LLM-Based Tutor Agent with a Large Action Model for Multimodal Interaction | DOI |
🔎 Peer reviewer for AAMAS '26 · CALM '25 / '26 · ArtInHCI '25 · MobiSPC '26 · EUSPN '26
|
EIF/I²TL grant-funded Socratic tutoring platform. Hybrid CrewAI multi-agent architecture, RAG over educator materials, adaptive learner profiling, and a live teacher dashboard with misconception tracking. Featured by FSTM.
|
Diagnostic RAG evaluation for Luxembourgish RC (640 questions, 16 texts). Compared BM25 variants, multilingual-E5, BGE-M3, and hybrid fusion across GPT-5.5, DeepSeek v4-pro, Claude Opus 4.8, and Sonnet 4.6. 197 automated tests. Paper submitted to LaTeLL 2026.
|
|
LoRA fine-tuned LuxemBERT, XLM-RoBERTa, and ModernBERT vs. multi-LLM pipelines across 11 diagnostic categories. Encoders outperform LLMs (38.7% macro-F1 vs. 30.7%). Full training on a 4GB GTX 1050 Ti in under 2 hours. Established a negative result on ensemble pseudo-labelling.
|
🇱🇺 LetzLearnAI-powered Luxembourgish learning app with 5 learning modes (conversation, grammar, vocab, quiz, immersion), SM-2 spaced repetition, ElevenLabs TTS/STT for pronunciation practice, RAG-grounded responses, and real-time LLM streaming.
|
🤖 AIA-PALMulti-agent system (LangGraph + CrewAI) that dynamically adapts course material and quizzes in real time based on student performance. Resulted in two peer-reviewed publications at AAMAS '25 and MobiSPC '25.
|
Self-hosted infrastructure on repurposed hardware — Docker, Portainer, NGINX Proxy Manager, AdGuard Home (network-wide DNS filtering), Tailscale mesh VPN, and n8n workflow automation. Sandbox for open-source AI tools and services.
|
Languages
AI / ML
Frontend
Backend & Data
Infrastructure & Tooling