Jeison Mesa
Statistician and data scientist. Time series, state-space models and machine learning in R and Python.
Summary
I am a statistician and data scientist with more than ten years of work in R and Python. My degree is in statistics, and my thesis compared estimators of the fractional differencing parameter in long-memory series. Since then the job has mostly been getting models out of a research setting and into production, which in practice has meant PySpark, Parquet, Docker and SQL alongside the statistics.
At CIAT I spent five years modelling how climate change affects crops and other productive sectors, using probabilistic time series and state-space models, and I worked on the statistical models behind the AClimate Colombia seasonal forecasts. As part of the CIAT climate services team I shared in the 2017 UNFCCC Momentum for Change award.
Since 2021 I have worked in industry on churn, segmentation, pricing, forecasting and recommendation for Globant’s clients. I led the team on subscriber segmentation for a media client. A churn model earns its keep when its drivers become customer segments that a marketing team can write a campaign for, so that is where I put the effort.
I have taught statistics at three universities and co-directed an undergraduate thesis. My publications sit mostly in climate and agriculture: seven journal articles and two reports as co-author. I currently work as an AI engineer at Globant, with Disney as the client, where I work on the agent design of LLM systems built with LangGraph and Azure OpenAI.
Experience
AI Engineer, Globant (client: Disney) Dec 2025 – Present · Remote, Colombia
- Work on agent design for multi-agent LLM systems, including prompt engineering and the contracts between agents.
- Designed and implemented a multi-agent system that reads a codebase and writes its documentation. Led the prompt engineering and set the split between deterministic and LLM steps, and added schema validation of agent outputs so that a bad output fails loudly instead of slipping through.
- Built a conversational intake agent for internal product requests. It pushes back on vague or solution-first input and flags overlap across teams.
- Built the ETL and LLM engine behind scheduled reports. Live data from project-management tools goes into one reporting model under extraction rules that keep field-level provenance, with pre-distribution checks before anything is sent out.
Tools: Python, LangGraph, Azure OpenAI, REST APIs, Docker, Git, CI/CD.
Data Scientist, Globant May 2021 – Dec 2025 · Remote, Colombia
Sports-industry client (Aug 2023 – Dec 2025)
- Built statistical and machine learning models for sports-industry products, mainly pricing and match analytics.
- Worked with time series, econometric methods and embeddings on historical sales and event data.
Augmented Staffing, internal to Globant (Feb 2022 – Jan 2023)
- Owned the end-to-end forecasting and clustering of internal staffing tickets, used to plan recruiting.
- Built a semantic search and recommendation tool that matched candidates to tickets on text embeddings, then added a seniority estimation model in PyTorch with an ordinal loss to improve the match. Pipelines ran on Kubeflow.
Media and telecommunications client (May 2021 – Dec 2021)
- Led the team on subscriber segmentation and directed the data science work on the churn model and the marketing campaigns.
- The churn model combined platform usage with sentiment from text, using custom neural networks in PyTorch. Its feature importance fed the segments.
- Retention and acquisition email campaigns drew on an econometric analysis of subscription pricing and a market basket analysis. Pipelines ran on AWS SageMaker, with data from Athena and S3.
Professor (part-time), Universidad del Valle and Universidad Católica Aug 2020 – Jun 2021 · Cali, Colombia
- Taught Time Series Analysis and Probability at Universidad del Valle, and Fundamentals of Statistics I and II at Universidad Católica.
Data Scientist, CIAT (International Center for Tropical Agriculture) Jan 2015 – Mar 2020 · Cali, Colombia
- Modelled climate variables and their effect on crops and other productive sectors with probabilistic time series models, state-space models and spatio-temporal stochastic processes.
- Developed the statistical models behind the AClimate Colombia seasonal forecast platform (paper).
- Studied near-term climate change impacts on major crops in Latin America and the Caribbean for the IDB and USAID. Set up multiple DSSAT and ORYZA runs from R for the region’s five main production systems and combined them with machine learning and geostatistics.
- Mapped climate service hotspots worldwide for USAID (2017–2018), meaning places where agro-climatic information could help close yield gaps. GAMs and other machine learning, the whole process in R (AgMetGaps).
- Built an index with spatial machine learning models to support the breeding of climate-smart crops.
- Orinoco regional climate change plan (Meta, Casanare, Vichada and Arauca, 2015–2016): climate clusters, geospatial models automated in R, and yield simulation for eleven production systems in the Llanos.
- Public R code I contributed to: usaid_forecast_maize and AgMetGaps.
Student Intern, CIAT Apr – Oct 2014 · Cali, Colombia
- Classified production systems with cluster analysis for Colombia’s Fondo Adaptación, with CORPOICA, as part of a vulnerability analysis for areas hit by the rainy season.
Earlier work 2010 – 2013
- Consultant, Universidad del Valle (2012–2013): statistical design and analysis for two engineering theses, one on constructed wetlands (factorial and repeated-measures designs, mixed models), one on cubic equations of state (non-linear models, bootstrap).
- Professor, Universidad Santiago de Cali (2013): statistics in R for research faculty.
- Statistical Assistant, Universidad del Valle (2010–2013): analysis and R automation for accreditation in the health faculty.
Skills
Text and LLMs: agents and multi-agent systems (LangGraph, LangChain), prompt engineering, embeddings and vector search, topic models, Hugging Face, OpenAI and Azure OpenAI, LangSmith, PromptLayer.
Time series and statistics: state-space models and Kalman filters, ARIMA, GARCH, long-memory models, Gaussian processes, regression and mixed models, GAMs, experimental design, bootstrap, geostatistics.
Machine learning: clustering, dimensionality reduction (PCA, kernel PCA, t-SNE, UMAP), gradient boosting (XGBoost), deep learning (PyTorch), anomaly detection, scikit-learn.
Programming: R (advanced), Python, SQL.
Cloud and data: PySpark, sparklyr, Polars, pandas, Parquet, raster and satellite data, Databricks, AWS (SageMaker, Athena, S3), GCP (Vertex AI, BigQuery), Azure (OpenAI, Functions, DevOps).
MLOps and engineering: Git, GitHub, GitLab, Docker, Kubeflow, CI/CD, DVC, Datadog, Linux, Quarto, R Markdown, LaTeX.
Domain: climate impact modelling, crop simulation (DSSAT, ORYZA), churn, customer segmentation, pricing.
Languages: Spanish (native), English (B2, upper intermediate).
Education
B.Sc. in Statistics, Universidad del Valle, Cali (2008 – 2014). Thesis: Comparison of estimation methods for the fractional differencing parameter in conditionally heteroscedastic long-memory time series.
Thesis supervision
Co-director of an undergraduate thesis in Statistics, Universidad del Valle (2023), Spectral analysis of Colombian genres: La Cumbia (in Spanish). Open access.
Talks
Principal speaker on the thesis at the ESPOL statistics and informatics conference (Ecuador, 2014) and at a Universidad Nacional de Colombia and Universidad del Valle summer school (2013). Co-author at the AGU Fall Meeting (2016).
Awards
UNFCCC Momentum for Change award, 2017, as part of the CIAT climate services team. Best Research Outcome, CIAT, 2017.
Selected publications
Citation counts are on Google Scholar. Full list on jeikosd.github.io/publications.
- Valencia JB, Guryanov VV, Mesa-Diez J, et al. (2024). Predictive assessment of climate change impact on water yield in the Meta River Basin, Colombia: an InVEST model application. Hydrology, 11(2), 25. DOI
- Valencia JB, Guryanov VV, Mesa-Diez J, et al. (2023). Assessing the effectiveness of the use of the InVEST annual water yield model for the rivers of Colombia. Water, 15(8), 1617. DOI
- Ordoñez L, Vallejo E, Amariles D, Mesa J, et al. (2022). Applying agroclimatic seasonal forecasts to improve rainfed maize agronomic management in Colombia. Climate Services, 28, 100333. DOI
- Ochoa-Muñoz AF, Peña-Torres JA, García-Bermúdez CE, Mosquera-Muñoz KF, Mesa-Diez J (2022). On characterization of sensory data in presence of missing values: the case of sensory coffee quality assessment. Ingeniare, 30(3), 564–573. DOI
- Valencia JB, Mesa J, León JG, Madriñán S, Cortés AJ (2020). Climate vulnerability assessment of the Espeletia complex on páramo sky islands in the northern Andes. Frontiers in Ecology and Evolution, 8, 565708. DOI
- Sotelo S, Guevara E, Llanos-Herrera L, et al., Mesa J, et al. (2020). Pronosticos AClimateColombia: a system for the provision of information for climate risk reduction in Colombia. Computers and Electronics in Agriculture, 174, 105486. DOI