Felipe Espinosa

• Architected a data analysis agent that achieved 88% end-to-end QA accuracy on an internal benchmark. • Reduced SQL drafting time from minutes to 51 seconds on average by implementing a self-correcting agent loop. • Optimized RAG pipeline retrieval to 95% Top-k accuracy by fusing BM25 and dense embeddings with Reciprocal Rank Fusion. • Integrated multi-tool capabilities (coding, data visualization, and analytical reasoning) to enable complex, autonomous analyst workflows. • Implemented active learning loops capturing user feedback for few-shot grounding and fine-tuning data curation. • Deployed the system to a pilot group within the Data team to validate the architecture for future enterprise integration.

• Trained the first Text-to-Speech model for Walloon (3h of data) using cross-lingual transfer from 20h of French, achieving MOS 4.22/5 for the best configuration. • Ranablationstudies (French pretraining vs none, graphemes vs phonemes) and conducted objective + human evaluation (PESQ, MCD, SECS + native-speaker perceptual tests) to select the best configuration. • Lead author of a paper presented at the 13th Speech Synthesis Workshop (SSW 2025), in collaboration with senior researchers at LISN, highlighting cross-lingual transfer strategies for low-resource TTS. • Open-sourced trained models, demos, and inference scripts on Hugging Face, enabling reuse by other researchers.

• Worked on residential and commercial renovation projects, delivering scoped tasks and supporting on-site execution. • Coordinated with clients, supervisors, and trades to meet schedules, requirements, and safety standards.