- 🧮 Mathematician & statistician with 20+ years turning data into business decisions in insurance, banking and healthcare.
- 🤖 Now building Generative AI systems end to end: RAG with embeddings and vector search, LangGraph agents, voice bots and computer vision.
- 📏 I measure what I build — retrieval metrics, statistical tests, honest limitations.
- 🌍 Based in Lugo, Spain · working remotely.
AI Portfolio – María Tocado Murillo ├── 🍳 CulinarIA – LLM Recipe Assistant LangGraph · RAG · FastAPI · voice ├── 🏛️ Odyssey Hybrid RAG (Master's thesis) Qdrant · hybrid retrieval · statistics ├── 🚗 Traffic Detection & Tracking YOLO11 · DeepSORT · computer vision ├── 💬 NLP: from TF-IDF to Transformers Word2Vec · LSTM · sentence-transformers ├── 👕 Fashion-MNIST CNN PyTorch · CNN · test accuracy 92% ├── 🔍 Sales Prediction · Data-Leakage Audit scikit-learn · CV · permutation test ├── 🏥 Hospital Readmission · 9-Model Comparison R · tidymodels · bootstrap CI ├── 📡 Churn Prediction API scikit-learn · FastAPI · pytest └── 🔒 Conversational AI · Voice Bot & CRM private · client project
LangGraph agent with LLM routing, conversational memory, RAG over Chroma and deterministic allergen safety. FastAPI backend + web UI with voice.
|
Hierarchical hybrid retrieval (dense + lexical, RRF) over Qdrant. Hit@1 = MRR = 1.00 on independent validation · RAG beats baseline (p = 0.027). Master's thesis, featured by UDIT.
|
Fine-tuned YOLO11s + DeepSORT on 52,851 driving images. Test mAP@0.5 = 0.776.
|
Sentiment analysis across four stages: TF-IDF, Word2Vec, LSTM in PyTorch and sentence-transformers.
|
CNN clothing classification with a clean train / validation / test protocol. Test accuracy 92% · T-shirt recall 0.93 · error analysis across three models.
|
A model with R² = 0.92 that had learned nothing: target leakage detected, removed and verified with repeated CV and a permutation test (p = 0.99).
|
Logistic, Lasso, tree, random forest, XGBoost, KNN, SVM and neural net, one protocol. Bootstrap 95% CIs show no model beats plain logistic regression.
|
A scikit-learn pipeline served with FastAPI — token auth, rate limiting, full validation. Two real bugs found and fixed; 12 tests passing.
|
Private · client project. Web / Instagram / WhatsApp assistant with voice (STT/TTS), LLMs, REST APIs and CRM automation.
|
Master's in Artificial Intelligence (UDIT) · Degree in Mathematical Sciences (UCM) · Azure AI Fundamentals
