I'm Íñigo Fernández Barrill, a software engineer from Pamplona, Spain, specialized in AI / Machine Learning and Full Stack development.
- 💼 I worked as a Full Stack Software Developer at Orisha Commerce (Openbravo), Mutilva (Navarra), from September 2023 to March 2026:
- Validated a three-agent LLM PoC fine-tuned with LoRA (customer interaction, database queries, RAG over POS docs); the discount and customer-interaction models went into production, with over 80% of interactions returning the expected result.
- Designed and built REST APIs integrating the Openbravo ERP with external e-commerce platforms, serving 3,500+ POS terminals.
- Cut production-reported errors in a critical ERP module by ~50% by introducing a unit/integration testing framework.
- 🔬 Before that, I was a research intern at Veridas Digital Authentication Solutions (Noáin, Navarra), across two periods between 2021 and 2022: improved OCR effectiveness by 12%, cut image-analysis processing time by 85%+ by automating manual processes, and turned 500,000+ files into high-quality training data.
- 🎓 Master's Degree in Data Science and Machine Learning — University of La Rioja (2022–2023). My thesis applied Multi-Task Learning to breast cancer detection in mammographies (F1 0.68 with only ~2% positive cases) →
CancerDetection. - 🎓 Computer Engineering Degree, specialization in Computing/AI — Public University of Navarre (2017–2022).
- 🌍 Spanish (native) · English (C1) · French (B1).
- 🚀 Currently looking for opportunities as a GenAI, Machine Learning, Data Science or Backend (Python) Engineer. Available immediately · Pamplona, Spain (CET) · open to remote. If you have something that could be a fit, let's talk!
AI / ML: PyTorch · FastAI · Timm · scikit-learn · Pandas · NumPy LLMs / GenAI: LoRA / QLoRA · Unsloth · vLLM · llama.cpp · Ollama · LangGraph · LangChain · LlamaIndex · RAG · MCP · Hugging Face · MLflow Backend / Full Stack: Python · Java · Node.js · FastAPI · Flask · JavaScript / TypeScript · HTML / CSS · React · Next.js · Expo / React Native Data & infrastructure: PostgreSQL · MySQL · MongoDB · SQLite · DuckDB · LanceDB · Docker · Git · GitHub Actions · Jenkins · Playwright / Scrapling
|
|
| Project | What it is | Stack |
|---|---|---|
| jobradar | Self-hosted job radar: searches job boards, scores every opening against your CV and writes a tailored one-page CV for each one, verified so nothing is invented. Restricted boards (LinkedIn, InfoJobs, Tecnoempleo, Indeed) are opt-in and read through a real browser. 470+ tests and CI; on PyPI as jobradar-cv. |
Python · FastAPI · SQLite · LLMs (Gemini, Anthropic, OpenAI, Ollama) · MCP · Playwright · Scrapling · Docker · GitHub Actions |
| navarra-trip-agent | Agent that plans multi-day trips around Navarre from the Government of Navarre's open data: a day-by-day plan with real travel times, a map, lodging and restaurants from the official tourism registry, and the source of every place, or a short list of options ("pintxos bars in Estella"). Evaluated on 93 requests in Spanish, English, French and Basque with code validators and a calibrated LLM judge, comparing five models: a 7B model running locally matches a 120B one (100% real places, 100% constraints met). Ships its own MCP server and a skill. Live demo on Vercel's free plan: FastAPI as a serverless function with Groq, multilingual-e5 in ONNX for semantic search, precomputed OSRM travel times and Upstash Redis. | Python · LangGraph · Groq · Ollama (Qwen2.5) · MCP · DuckDB · ONNX Runtime · OSRM · FastAPI · Next.js · TypeScript · MapLibre · Vercel · Redis |
| boe-extractor | Turns documents from the BOE (Spain's official gazette) into schema-validated JSON with Qwen2.5-1.5B fine-tuned with QLoRA. Quantized to 4 bits, it runs on CPU and performs on par with the paid API model it learned from (F1 0.92 vs 0.91 for Gemini Flash-Lite on hand-annotated documents; base model 0.66, rule-based baseline 0.85), and cuts hallucinations from 12% to about 1%. Evaluated on 262 reviewed documents, tracked in MLflow. A daily GitHub Actions job publishes the day's public-sector jobs, grants and tenders on a web page. Model and dataset on Hugging Face. | Python · Unsloth · QLoRA · vLLM · llama.cpp (GGUF) · Pydantic · MLflow · FastAPI · GitHub Actions · Hugging Face |
| gym-coach | Offline mobile app (Android and iOS) to log gym training: routines with schedules and reminders, progress charts, technique cards with a muscle map and level per muscle group. 56 exercises with 179 variants and custom illustrations. GitHub Actions builds the Android APK and publishes it as a release. | TypeScript · Expo · React Native · expo-router · SQLite · Drizzle ORM · Jest · GitHub Actions |
| CancerDetection | Master's thesis: Multi-Task Learning for Breast Cancer Detection in Mammographies. Compares image models (ResNet, EfficientNet, ConvNeXt, HRNet, DeiT), classical tabular models and two Multi-Task Learning architectures on the RSNA Screening Mammography (Kaggle) dataset; best F1 0.68 with only ~2% positive cases, through rebalancing and feature fusion. | PyTorch · FastAI · Timm · scikit-learn |
Live counts:
Thanks for stopping by 🙂 — let's connect on LinkedIn