Trujillo, La Libertad, Peru
I build and debug data pipelines around LLMs — retrieval systems, large-scale web scraping, and LLM-based classification. What I'm good at is diagnosis: taking a system that returns "wrong" results, isolating the actual cause, and proving the fix with numbers rather than guessing. On a recent search project over 3.5M documents, I showed the bottleneck was data preparation and index architecture — not the embedding model — and moved on-topic results from near zero to 93%. I work in Python, lean heavily on measurable evaluation (gold sets, Precision@k, MRR), and treat working alongside AI coding assistants as a core part of how I ship. Based in Peru, open to remote and contract work.
Designed and maintained automation systems for data processing and social-media operations under a service contract. Built automated social-media posting and analytics-collection pipelines, and developed a Telegram bot serving as the prototype for a future content management system. Responsible for keeping the systems running, supervising updates, and fixing errors as they arise.
Experience — Content search platform (RAG / retrieval): AI Automation Specialist Diagnosed and improved a retrieval system powering search over ~3.5M articles. Built a gold-set evaluation (Precision@5, MRR) and proved the quality bottleneck was data preparation and centroid architecture, not the embedding model. Shipped four targeted fixes that raised on-topic results from ≈0 to 93%, improved P@5 by 23%, and increased niche accuracy 2.5×. Designed a custom RU-web taxonomy (~30 topics + subtopics) replacing the standard IAB model for better corpus fit. Built a data-enrichment pipeline: LLM classification + filtering, benchmarked across 5 models, with batching for a 12× speedup, indexed into Qdrant. Deployed backend + frontend in Docker and validated the service end-to-end.