I build production systems across three fronts — agentic AI (LangGraph, RAG, multi-agent), NLP and text analytics (semantic search, embeddings, bibliometric networks), and applied data science (clustering, dashboards, pipelines) — for research and product teams across Latin America.
I'm an economist and computer-science engineer (M.Sc.) working across three fronts: agentic AI (LLM orchestration, RAG, multi-agent systems), NLP and text analytics (semantic search, embeddings, bibliometric networks), and applied data science (clustering, statistical modeling, dashboards). My path moves between them — sometimes within a single project.
That breadth means I do the technical work and understand the social, organizational, and research context behind it. I've shipped conversational agents serving hundreds of users monthly across Latin America, ML clustering pipelines for survey research, RAG systems with vector search, and 6 monitoring dashboards across 4 countries for data-capture and impact-evaluation processes.
Currently leading data science and AI at Estudio Plural — designing LLM-based tools for behavioral research, knowledge retrieval, and organizational intelligence. I publish peer-reviewed work on bibliometric NLP, teach, and consult on applied research projects when there's a good fit.
Production systems across three fronts — agentic AI, NLP, and data science — built for research and product teams.
Interactive map of Cali's 339 neighborhoods with the 2026 first-round presidential results. Official Registraduría tally (216 polling stations, 5,158 tables) geolocated to neighborhood level by cross-referencing IDESC's WFS school layer + OpenStreetMap geocoding. Leaflet, static site on Vercel.
Social-listening platform mapping the public X/Twitter conversation around Colombia's 2026 presidential race. Apify ingestion → LLM classification of emotion, framing and stance → embedding-based narrative clustering (fastembed + HDBSCAN) → a directed interaction graph with coalition alignment. Next.js 16 static site on Vercel.
Digital social-listening pilot on the Spanish-speaking online conversation (TikTok focus), built for an Estudio Plural × Camino proposal. 742 unique videos over four months, each AI-classified as signal vs noise — only ~35% of keyword matches are genuine, so raw volume overstates the topic roughly 3×. Hashtag-and-creator network, geography, engagement by theme, and a video-by-video corpus explorer. Next.js on Vercel.
Interactive dashboard for Colombia's General Royalties System (SGR). Real-time data from datos.gov.co via Socrata API, dynamic filters, choropleth maps, and Excel export. Deployed on Streamlit Cloud.
Survey-processing pipeline for AMA's social field research across multiple cities. KoboToolbox ingestion, validation and deduplication feeding an interactive endline results dashboard — auto-generated charts, cross-tabs, and LLM-written report sections. Deployed on Streamlit Cloud.
Monitoring dashboard for the AMA WhatsApp bot. Tracks user activity, sessions, and engagement across deployments. Streamlit + Supabase backend with Plotly visualizations and Excel exports.
End-to-end ML pipeline for archetype discovery. LangGraph orchestrates ingestion → profiling → preprocessing → algorithm selection → clustering → LLM-generated narrative. 33 automated tests passing.
Multilingual bot (ES/EN/PT) for Equimundo's A+P Manual. 5 sequential LLM agents: language detection → intent classification → specialized response (factual, planning, ideation, sensitive topics). Built with FastAPI + LangGraph.
Automated daily scanner of 15+ funding and grant sources. Claude AI filters by organizational relevance, deduplicates results, and sends curated alerts to Slack. Runs on GitHub Actions every morning.
Operational monitoring dashboard for the Aly (Apapáchar) WhatsApp bot. KPIs with sparklines and deltas, geographic visualization, alert flags with Excel export and review-status toggle, and a leaderboard with drill-down. Multi-page Streamlit app with custom navigation and i18n.
Field-survey validation pipeline for the AMA program in Leticia (Colombia) and Cobija (Bolivia). KoboToolbox QC, ID validation, duration outlier detection per classroom, attendance crosschecks vs Google Forms, school-level Excel reports, and LLM-generated narrative summaries via OpenRouter.
No-code SaaS platform for building multi-agent chatbots with multi-channel deployment (WhatsApp, Telegram, Web). Full UI in Next.js + shadcn/ui; FastAPI backend with MongoDB Atlas and Supabase auth.
datos.gov.co Open Data API → 8 hours/week saved for the project formulation team.Bibliometric NLP and citation-network analysis applied to economic discourse — 104K+ nodes across four peer-reviewed studies.
Hire me to take an idea from prototype to production — across conversational AI, data systems, and applied research.
Production chatbots and multi-agent assistants — multilingual, with RAG over your documents and conversation memory. Deployed where your users already are.
From messy field data to decisions your team can act on: ingestion, automated validation and QC, and live dashboards they'll actually use.
Research-grade text and data analysis — semantic search, classification, sentiment, clustering and network analysis over surveys, documents and organizational data.
Open to consulting, research collaborations, and new projects — especially where AI, data, and social impact intersect.