At CTCAP Group, I led architecture and core technology decisions for LAI, a legal due-diligence AI platform for German wind-energy projects. I designed and tuned hybrid retrieval over a 35M+ passage legal corpus, with citation verification and an evaluation harness.
I also built LAI’s document-processing and LLM-serving stack, including German legal PDF/OCR cleaning and self-hosted Qwen3 inference via vLLM. I coordinated a three-person core team and added quality checks to catch regressions before release.
For Jarvis, I built a self-hosted voice and text assistant that runs on CPU without cloud APIs. Its persistent RAG memory uses ChromaDB, while whisper.cpp and Piper provide speech recognition and text-to-speech.
In my federated-learning capstone, I compared FedBN, FedAvg, local-only, and centralized approaches to brain-tumor segmentation on BraTS 2021. I reported both the 2D results and the negative 3D finding, documenting which hypotheses the study supported.
