_{English · Español}

Junior AI Engineer · Building LLM-powered applications on Azure
_{Based in Madrid · Open to roles across Europe (remote or relocation)}

I build production-grade LLM applications — RAG systems, fine-tuned models, and AI-powered backends — with a strong bias toward retrieval-first architectures that don't hallucinate.

I'm completing a Professional Master's in AI and Data Engineering at Tajamar (Madrid) and hold four Azure certifications: AI-102, DP-100, DP-300, and DP-900. My most recent role was an AI Application Developer internship at Datarmony, where I built a financial statement ingestion app on Google Cloud + Gemini.

Most of what I build sits on the same stack: Azure AI Foundry + Azure AI Search + FastAPI + React, orchestrated with Python and the OpenAI SDK.

Full-stack multi-tenant RAG platform with structural index isolation (one Azure AI Search index per assistant), conversational memory, LLM-based query rewriting, and verifiable citations. Hard "I don't know" fallback when retrieval is empty — no hallucinations by design. ~2.6k LOC Python + ~1.5k LOC TypeScript, 56 unit tests, built in 7 days.

Azure AI Foundry · Azure AI Search · FastAPI · React · TypeScript · Hybrid Search · Semantic Reranking

Unsupervised credit-card fraud detection via reconstruction-error autoencoder. Chosen knowing supervised models score higher — matches the realistic production scenario where fraud labels lag fraud patterns. PR-AUC ~3× the Isolation Forest baseline, bracketed by a Logistic Regression upper bound. Exported to ONNX with <1e-5 numerical verification and served from the same model file in the browser. Data-leakage prevention enforced via in-code assertions.

Deep Learning · Anomaly Detection · PyTorch · ONNX · Unsupervised Learning · TensorFlow.js · React

Real-time facial-emotion analyzer for sales calls. Two CNN architectures benchmarked end-to-end (custom 4-block CNN vs fine-tuned MobileNetV2, 63 % test accuracy) feeding a rolling receptivity index as a coaching signal. Live in the browser via TensorFlow.js — no install, runs on phone or laptop. FastAPI service + Streamlit demo share one inference module. 4 pedagogical notebooks documenting every training failure and fix.

Deep Learning · Computer Vision · TensorFlow · TensorFlow.js · CNN · Transfer Learning · FastAPI · Streamlit

Full-stack SaaS for Spanish freelancers and SMEs. Uploads invoice PDFs, runs structured GPT-4o extraction via Azure AI Foundry, validates the JSON, and surfaces a dashboard with KPIs, monthly trends, top clients/suppliers, and quarterly VAT. Built in one week with a teammate. Deliberate single-call extraction (no agents) — more reliable, cheaper, and easier to debug for structured invoice data.

Azure AI Foundry · GPT-4o · Structured Extraction · React · TypeScript · Node.js · Express · SQLite

Fine-tuned gpt-4o-mini to act as a financial education assistant with consistent format, tone, and legal disclaimers. Custom JSONL dataset, overfitting analysis, multi-client behavior debugging.

Fine-tuning · Azure AI Foundry · gpt-4o-mini · JSONL · Evaluation

Two-level transaction categorisation pipeline for a Spanish fintech aggregator. LightGBM L1 (91.2% accuracy) escalates to Azure OpenAI gpt-4o-mini only when confidence drops below 0.70 — combined accuracy 96.1%. GDPR-safe by design: Presidio + regex anonymisation strips PII before any external call. 3,479 LOC Python, 55 tests passing. I owned the architecture, cost-tier strategy, GDPR pipeline, and full app engineering (API, anonymisation, anomaly detection, insights, dashboard); my teammate trained the ML classifier.

LightGBM · Azure OpenAI · Presidio · GDPR · FastAPI · Streamlit · Anomaly Detection

AI / ML

Azure

Backend

Frontend

Cloud & Tooling

• Microsoft Certified: Azure AI Engineer Associate (AI-102)
• Microsoft Certified: Azure Data Scientist Associate (DP-100)
• Microsoft Certified: Azure Database Administrator Associate (DP-300)
• Microsoft Certified: Azure Data Fundamentals (DP-900)
• Professional Master's in AI and Data Engineering — Tajamar, Madrid (2025–)
• Web Application Development — IES Pío Baroja, Madrid (2023–2025)
• Cambridge English: C1 Advanced

TFM @ Integra — Public tender analysis system (team of 3, private repo) Building an AI-powered system that parses public tender documents (licitaciones), extracts structured data, summarizes them in plain language, and helps the company draft stronger bid proposals. My role: backend + AI implementation on Azure AI Foundry with a RAG pipeline over tender corpora.

RAG · Azure AI Foundry · Document Analysis · FastAPI · Team project

Also:

• Designing a multi-agent system for automated job application management
• Open to Junior AI Engineer roles — Europe-wide, remote or relocation. Based in Madrid.

Hiring a Junior AI Engineer? I'm based in Madrid and available for full-time roles across Europe — remote or relocation. I build production LLM and ML systems on Azure that solve real business problems, not just demos. Reach me by email or LinkedIn — I reply within 24h.