Work
Projects
A selection of research projects, production deployments, and engineering work.
Care AI: Multilingual AI for Dementia Care AI4Good winner
Multilingual AI communication assistant helping caregivers in German care homes connect with migrant residents in their native language. As dementia progresses, residents often revert to their mother tongue and the cultural world of their youth; generic translation fails here. The system uses Validation Therapy (the evidence-based standard in dementia care) to deliver situation-aware emotional guidance: what to say in the resident's language, a concrete calming action grounded in their life story, a possible reason for the behaviour, and cultural context.
Designed and built full-stack: React Native PWA, FastAPI backend, LLM guidance pipeline (Grok), Whisper STT, edge-tts for native-language speech, and Tesseract OCR for medical document scanning (German discharge letters to other languages). Each resident has a personalised memory profile: life story, languages, calming factors, triggers, photo bank, which the AI references by name in its guidance.
The core technical differentiator is an in-context RL feedback loop: caregiver ratings (helpful / not helpful) are stored with situation context and scored across four signals: language match, mood match, time-of-day, and keyword/situation overlap (Jaccard). Top-ranked validated responses are injected as few-shot examples in subsequent prompts, alongside a negative block of unhelpful patterns. The system improves with use, per resident, with no model retraining.
AI4Good Hackathon · 2026
ENVELOPE: Deep RL for Multipath 5G Traffic Steering
Research software development on the ENVELOPE project at Leibniz Universität Hannover, focused on modelling Access Traffic Steering, Switching and Splitting (ATSSS) with Multipath QUIC (MPQUIC) as a reinforcement learning environment, implemented using RLlib. The work required a solid understanding of RL algorithms, multipath protocols, and network traffic dynamics.
Conducted experiments on a real-world 5G testbed and analysed the developed approach under varying network conditions. Built a visualisation tool that automatically generated training and evaluation plots, and developed experiment orchestration and testbed automation tooling that reduced manual effort and improved reproducibility.
Co-authoring: Multi-Access Traffic Steering in Real-World 5G with Deep Reinforcement Learning (paper in preparation).
Leibniz Universität Hannover · Apr 2025 – May 2026 · Research assistant
Production RAG & LLM System: Mental Health NLP
Led a team of 30+ engineers at Omdena to build and deploy a production LLM + RAG pipeline in 9 weeks from prototype to live AWS deployment. The system processed unstructured, culturally varied mental health and wellbeing data with significant noise: inconsistent language, missing fields, cross-cultural variation.
Diagnosed production data quality failures and led implementation of a user feedback loop (RLHF) to close the test/production performance gap. Designed evaluation loops, structured output schemas, and validation logic to reliably extract signal from subjective text. Owned the full pipeline: ingestion, embeddings, model integration, API deployment, and stakeholder delivery.
Omdena · Feb – Apr 2024 · Remote
Intelligent Financial Planning System
Built an LLM + RAG-based financial planning system using LangChain agent workflows, deployed on AWS with full API integration. Implemented structured output validation and fallback logic to ensure production-grade reliability of AI responses.
OptimumAI Labs · Apr – Jun 2024 · Remote
GPS Sensor Fusion Pipeline
Multi-sensor data fusion pipeline for real-time GPS localisation using Kalman filtering. Fused inconsistent signals from multiple GPS sources into a single reliable state estimate, handling measurement noise and sensor dropouts.
Universität Duisburg-Essen · Jun – Jul 2020
Autonomous 3-DOF Greenhouse Robot
Developed an autonomous robot system for greenhouse plant care with real-time sensor integration, motion planning, and control. The robot performed autonomous navigation and plant maintenance tasks without human intervention.
Universität Duisburg-Essen · Nov 2020 – Feb 2021
Neural Network System Identification & NMPC
MSc thesis: benchmarked NARX, LSTM, Bi-LSTM, and GRU architectures for nonlinear system identification, recovering reliable behavioural models from noisy, incomplete observations. Implemented a data-driven Nonlinear Model Predictive Controller (NMPC) and validated it in MATLAB.
Universität Duisburg-Essen · 2022 · MSc Thesis