Felipe Espinosa Orjuela

Felipe
Espinosa Orjuela

AI/ML Engineer

LLM agents, generative models and evaluation

Paris, France. Authorised to work in France, available now.

Portrait of Felipe Espinosa Orjuela

Selected work

2025, Capital Fund Management, Paris

An LLM agent that answers analysts' questions in SQL

Seven tools, an internal Oracle database, and a 48-question benchmark I built to measure it. Accuracy rose from 83% to 87.5%, and routing sub-tasks to smaller models made each query about 20× cheaper than o3.

Read the case studyInternship report (PDF)

Cost per query, at similar accuracy

o3 on the whole task
€5
The agent, small models on sub-tasks
€0.24
Approximate figures. Table retrieval reached 95% top-k with hybrid search.

2024, LISN, CNRS and Université Paris-Saclay

The first speech synthesis system for Walloon

VITS models trained on 18 to 156 minutes of speech. With only 18 minutes of the female voice, fine-tuning from French raised the listener score from 3.38 to 4.10 out of 5. First author, oral presentation at SSW 2025.

Read the case studyPaperTry the demo

Trained on Walloon only
From scratch, 13 s
3.38MOS
Fine-tuned from French
Same 18 minutes, 13 s
4.10MOS
Female voice, grapheme input. Mean opinion score out of 5 from 23 listeners; natural speech scored 4.55.

2026, personal project, in progress

AskData: plain-English analytics on PostgreSQL

A FastAPI and Next.js app that turns questions into read-only SQL, with schema-aware retrieval and automatic repair of failed queries. Benchmark results will be published here when the evaluation run is done.

Project notesCode

AskData answering 'Which 10 customers spent the most in total?' with a sentence and a bar chart

Also: a synthetic French NER dataset for clinical trials, course project with a team of three; 12,423 rows released on Hugging Face.Code

Research

  1. Under reviewCross-lingual study of phoneme, syllable and word embeddings in Spanish, English and French. Lead author. Language Resources and Evaluation (Springer Nature).
  2. 2025Speech synthesis for Walloon, an under-resourced minority language. Espinosa Orjuela, Boula de Mareüil, Evrard. 13th ISCA Speech Synthesis Workshop (SSW 2025), oral. Poster at LT4All 2025 (UNESCO).
White four-wheeled rover with USP and Embrapa logos in a maize field, a drone resting on top

2014, University of São Paulo, Mobile Robotics Lab

Propulsion and steering for an agricultural rover

Built with Embrapa Instrumentação. I selected the motors, controllers and sensors, wrote the motion and collision-avoidance routines, and ran the field trials.

  1. Projeto mecatrônico de um rover para análise de solos usando tecnologia LIBS, Parts I and II. Co-author. SIAGRO 2014, Embrapa Instrumentação.
  2. Controle da propulsão e o esterçamento de um rover aplicado à agricultura de precisão. Co-author. SIICUSP 2014, University of São Paulo.

At SIAGRO 2014; the drone was a hand-flown demonstrator. Photo: Tati Zanichelli.

Experience and education

Experience

  1. Research Collaborator, NLP, Universidad Tecnológica de Bolívar2026 – now
  2. AI Evaluation Contractor, Mercor, Mindrift, Alignerr, DataAnnotation2025 – 2026
  3. AI Engineer Intern, LLM Systems, Capital Fund Management2025
  4. AI Research Intern, Speech Synthesis, LISN, CNRS2024
  5. Self-Employed Contractor, construction, own business, Sydney2018 – 2023
  6. Research Assistant (exchange), University of São Paulo2014

Education

  1. M.Sc. Artificial IntelligenceUniversité Paris-Saclay. 16.2/20, Mention Très Bien2023 – 2025
  2. Advanced Diploma of Leadership and ManagementLloyds International College, Sydney2020 – 2022
  3. B.Sc. Electronic EngineeringUniversidad Industrial de Santander. GPA 3.89/52010 – 2016

Tools. Python, PyTorch, Hugging Face, LangGraph, SQL, C/C++, Docker, Slurm.
Languages. Spanish, English (C2), Portuguese, French (intermediate).

Contact, for roles and CV requests

LinkedInGitHubGoogle ScholarHugging Face