AI/ML Engineer in Berlin · Applied AI systems with a research foundation — I build LLM/ML systems that ship, and I know when not to use AI.

🇹🇼 Taiwan-born · 🇩🇪 Berlin-based · Full unrestricted work authorization in Germany 🗣️ Chinese (native) · English (C1) · German (B2→C1) · French (B2)

📄 LinkedIn · 📚 Published research

Electricity Price Forecasting End-to-end MLOps pipeline: LightGBM/LSTM/XGBoost benchmark → Prefect 3 orchestration → MLflow tracking → FastAPI serving → Docker/Kubernetes → Prometheus/Grafana monitoring → GitHub Actions CI/CD. LightGBM Test MAE: 7.23 EUR/MWh (best of benchmark) MLOps LightGBM Kubernetes Prefect MLflow

AWS AI Agent RAG + Text-to-SQL + LLM router with rigorous evaluation via RAGAS. Faithfulness 93.6% · Answer relevancy 92.7% RAG LLM Routing ChromaDB RAGAS AWS

XRCC Hackathon — Louvre XR Backend 🏆 2nd Place, WebXR (Meta IWSDK) Track FastAPI backend for a museum XR companion: GPT-4o Vision exhibit recognition, FAISS RAG, navigation lookup tables, multi-turn conversation — deployed on Railway. Built the case for deterministic lookup over GenAI where determinism wins (exhibit ID matching, 56-route navigation). Also shipped a same-weekend Android app integrating Ray-Ban Meta AI Glasses via the Meta Wearables DAT SDK, after discovering WebXR lacks camera/mic access on that platform. FastAPI GPT-4o Vision FAISS Android DAT SDK

MSc Thesis — Published in Mathematics (2025) A Deep Learning Framework of Super Resolution for License Plate Recognition in Surveillance Systems Built a real (not synthetically degraded) dashcam dataset; designed two novel perceptual losses (Swin Transformer– and CRNN-based) so the model optimizes for downstream OCR accuracy, not just pixel-level image quality. +10pp OCR accuracy improvement over baseline Swin Transformer CRNN Super-Resolution Perceptual Loss

Other applied ML/GenAI work:

• AI Brand Visibility Tracker — spaCy NER, LightGBM, GCP/BigQuery, RAGAS
• Churn Prediction — CatBoost, Optuna, Docker Compose, Dash
• SD Upscaler on Azure — FastAPI + Docker + ACI, CI/CD for AMD64 builds
• AI Parenting Chatbot — direct OpenAI API integration, custom similarity search
• Electricity Hyperband 🏆 3rd Place, Daytona Hackathon — Hyperband search using Daytona sandboxes + LightGBM

Hands-on implementations across discriminative, generative, and sequential modeling: Diabetic Retinopathy Detection (ResNet) · Conditional DDPM · Conditional VAE for Video Prediction · EEG Classification

💬 Open to AI Engineer / ML Engineer / MLOps Engineer roles — Berlin or fully remote (Germany/EU)