JOHN HANDLEY DERECHO
AI Engineer
john@johnhandleyd.com  ·  +34 670 087 316  ·  johnhandleyd.com  ·  linkedin.com/in/john-handley  ·  github.com/johnhandleyd
AI engineer experienced in client-facing delivery, with a track record in production LLM systems, agentic AI, and Retrieval-Augmented Generation (RAG). Hands-on Python developer comfortable owning architecture end-to-end.
Languages:  Spanish (native)  ·  English (C2)  ·  Swedish (A1)  ·  French (A2)
AI & Data Engineer  — Minsait (Indra) · Madrid
Mar 2026 – Present
Conversational voice AI agent  ·  Python (Google API); Dialogflow, ElevenLabs
• Lead engineer developing a voice AI agent in two environments: Dialogflow CX & Google Cloud, and ElevenLabs & ElevenAgents
• Handling team coordination as well as stakeholder expectations; collaborating with solution architect on deployment and cloud infrastructure
AI & Data Engineer  — EY wavespace · Madrid
Apr 2024 – Mar 2026
Graph-RAG Platform  ·  Python (LangGraph, Neo4j, Qdrant, Docker)
• Lead engineer on an enterprise Graph-RAG platform; boosted answer accuracy from ~30% to ~90% in 2 months through hybrid KG + embedding retrieval, improved query parsing and dynamic Cypher generation
• Designed a modular LangGraph multi-agent architecture with schema-aware Neo4j resolution, fuzzy/lexical matching and deterministic + LLM tooling; fully containerised with Docker
AI Agents — Virtual Assistants  ·  Python, JS (LangChain, CrewAI, FastAPI)
• Built production semantic-search assistants for client ticket retrieval, cutting lookup time from 5–10 min to seconds and saving 100+ hours/month
Programming: Python (OOP, async, APIs), Bash/Linux, Git, CI/CD, architecture prototyping
AI & ML: LLMs (OpenAI, Ollama), LangGraph, CrewAI, RAG, NLP (spaCy, HuggingFace), Dialogflow CX, scikit-learn, PyTorch, MLFlow, LLM-as-a-Judge, MCP
Data & Storage: Neo4j, Cosmos DB, PostgreSQL, SQL, ETL pipelines, vector search (Qdrant, Azure AI Search)
Cloud & Infra: Azure (DevOps, Key Vault, AI Search, Blob), Docker, Google Cloud, Agile (Scrum, Jira); Linux
Certifications: Azure AI Fundamentals (AI-900)
Antarctic Sea Ice Concentration from Sentinel-2
Python  ·  PyTorch  ·  rasterio
Weakly supervised regression (fine-tuned ResNet18): trained against coarse AMSR2 sea ice concentration labels (12.5 km) to predict per-patch SIC from Sentinel-2 L2A imagery at 10 m (2.56 km patches), resolving sub-grid-cell ice structure beyond AMSR2's native resolution.
Master's Thesis: Trained transformer models on 1.8M+ Wikipedia segments to analyse the effect of training data on neural machine translation quality
BSc Computer Science (in progress)  ·  Universitat Oberta de Catalunya  ·  Online2026 – Present
MSc Language Analysis & Processing (spec. Machine Translation)  ·  Universidad del País Vasco2020 – 2022
BSc Translation & Interpreting ·  Universidad Complutense de Madrid2016 – 2020