Intro
Hi, world. I’m Federico, an ML/AI engineer at Sabio Group. Previously, I worked at Overwatch AI, an aviation startup, and led multimodal NLP research at both the University of Colorado Boulder’s NSF AI Institute and the Spanish National Research Council (CSIC).
My area of expertise is NLP, and my personal favs are post-training, agentic workflows, and evaluation of LLM-powered systems. I also enjoy doing research now and then, especially on explainable AI and computational semantics.
In my free time, you’ll find me baking something, reading Agatha Christie, or learning a new language :)
Feel free to contact me at federico.ortegariba@fulbrightmail.org
Work Experience
→ AI Engineer @ Sabio Group | Aug 2026 – Present
Spain, Remote
-
Built real-time voice agents for the financial sector, orchestrating Deepgram (STT), ElevenLabs (TTS) and Twilio telephony, and cut p50 TTFA from 2s to 300ms.
-
Deployed self-hosted open-weight LLMs (DeepSeek-V4, GLM-5.2) to production with vLLM on GKE (GCP), backed by AWS S3 for storage.
-
Developed guardrails and evaluation frameworks for text and voice agents with Presidio and NeuralTrust for safety, and MLflow for experiment tracking.
→ AI Engineer (Contractor) @ Overwatch AI | May 2025 – Jul 2026
Spain, Remote
-
Designed and orchestrated multi-agent LangGraph systems for a pre-seed startup, fetching live data from external APIs via custom MCP servers and coordinating specialized agents over A2A.
-
Built a production RAG system with embeddings in Qdrant, deployed on AWS via Terraform, powered by closed- and open-weight models (Claude, GPT, Gemini, GLM-5.2, Nemotron-Ultra, Llama).
-
Established LLM evaluation (DeepEval, Ragas, TruLens) and OpenTelemetry observability via Phoenix (Arize), cutting p50 latency from 12s to 7s and tokens from 12k to 8k.
→ AI/Machine Learning Engineer @ University of Colorado Boulder | Feb 2025 – Apr 2026
Colorado, USA
-
Designed and deployed end-to-end retrieval pipelines (LangChain/LlamaIndex + Azure OpenAI) with vector search, improving Recall@5 from 85% to 90% and answer accuracy by 7 pts via chunking and evaluation.
-
Built scalable PyTorch/Hugging Face models for tool-use agents with RLHF-based reward modeling, improving task success rate from 74% to 85% on a 400-task evaluation set.
-
Refactored the production LLM serving codebase (Jigsaw Agent + Mistral) for the NSF National AI Institute for Student-AI Teaming, cutting inference cost by 40% via modular prompts and Langfuse observability.
→ AI/Machine Learning Engineer @ Spanish National Research Council (CSIC) | Oct 2023 – Apr 2024
Madrid, Spain
-
Fine-tuned BERT and T5 transformer models for domain-specific entity extraction on a 162k-token dataset, achieving F1 = 79% on held-out evaluation.
-
Built an offline evaluation framework on 500 curated test samples to identify top failure modes, prioritize iteration cycles, and surface ranking errors.
→ Machine Learning/NLP Engineer Intern @ University of Granada | Nov 2022 – Jun 2023
Granada, Spain
-
Designed a bilingual EN–ES dataset (300 entries) and built a data pipeline over a 600+ document corpus, extracting 150 domain keywords via automated tooling.
Education
→ MS in Computer Science (Natural Language Processing) @ University of Colorado Boulder | Aug 2024 – May 2026
USA
-
GPA: 4.0; fully funded by the Fulbright Scholarship (USDOS).
-
Coursework: Machine Learning; Deep Learning & Neural Networks; Advanced NLP.
→ MA in Computational Linguistics @ Complutense University of Madrid | Sep 2023 – Jul 2024
Spain
-
GPA: 3.5.
-
Coursework: Language & Speech Technologies; Advanced Programming; Statistics with R.
Publications
MY CS ESSENTIALS
A collection that will grow as I remember what I’ve read or keep reading new stuff:
SOME COOL PEOPLE & THEIR COOL STUFF
→ LLM design, evaluation & analysis