Machine Learning Engineer · Speech Recognition, NLP & LLM Agents · Spanish & Iberian languages
I build speech and language AI that works for real Spanish speakers, including the dialects and regional languages most systems get wrong.
- Balidea: next-generation Galician ASR for Proxecto Nós (~750 h; Whisper, Qwen3-ASR, Parakeet) and an on-premise LLM agent platform for the Galician Health Service, with RAG, tool calling and a safety layer.
- Universidade de Vigo: Ph.D. on parameter-efficient fine-tuning of audio transformers for clinical screening from voice (defense by April 2027).
- 2nd of 6 teams in both ASR and speaker diarization at ALBAYZIN 2026, on spontaneous, dialectal rural Spanish.
- R-LoRA: a LoRA variant with +12.4 pp Macro-F1 over standard LoRA on 4 clinical speech corpora (IberSPEECH 2026).
- First author in Scientific Data (Nature Portfolio, 2026): VisIA adolescent suicide-risk dataset.
| Repository | What it shows |
|---|---|
| clinical-safeguards | Fail-closed FastAPI guard middleware for clinical LLM apps; ~240 pytest tests, CI with a 90% coverage gate |
| ai-tutorials | Six bilingual (EN/ES) runnable notebooks: R-LoRA from scratch, Whisper + Optuna + MLflow, PEFT on Whisper, LLM guard internals |
| visia_q_dataset | Reproducible pipeline behind the Scientific Data paper (uv, pytest, CI) |
| acoustic_feat_extractor | Modular PyTorch extractor for acoustic features and speech embeddings |
| my-skill-hub | Custom agent skills for Claude Code (git, project-management and docs automation) |
Python · PyTorch · Hugging Face · Whisper · pyannote · Kaldi · LangChain · FastMCP · Milvus · FastAPI · pytest · uv · Hydra · Optuna · MLflow · Docker
Website · Hugging Face · LinkedIn · Google Scholar · ORCID · ramirezsanchezjosem@gmail.com