Blas M. Benito

@blasbenito.com

2395 Followers 2696 Following

🌍 Senior Data Scientist | Team Lead | 🌱 AgTech | PhD in Computational Ecology + MSc in GIS #rstats developer | geospatial engineering | soil microbiome and crop health mapping | product development https://github.com/BlasBenito - www.blasbenito.com

Jinshi, from The Apothecary Diaries, finding out something he never wanted to know.
Two pairs of kettlebells of 12 and 16 kilos.
RGB composite of Andalusia showing summer temperatures in red, normalized difference vegetation index in green, and annual rainfall in blue.
A linkedin job advert for a data scientist with a very particular requirement: the candidate must be a kaggle grandmaster.
igure 1: Overview of results: (Left) We find a significant decrease in library-specific skills (conceptual understanding, code reading, and debugging) among workers using AI assistance for completing tasks with a new python library. (Right) We categorize AI usage patterns and found three high skill development patterns where participants stay cognitively engaged when using AI assistance.
Global scatter plot of geographic point locations shown on a world map using longitude (x-axis) and latitude (y-axis). Blue hollow circles labeled “Training” are concentrated mostly in the Southern Hemisphere, covering South America, southern Africa, Australia, and parts of Southeast Asia. Red hollow circles labeled “Testing” are concentrated mainly in the Northern Hemisphere, covering North America, Europe, northern Africa, and large parts of Asia. The two datasets are largely separated by latitude, with minimal overlap around the equatorial region.
Two-panel world map illustrating spatial thinning. The top panel shows a very dense global distribution of points clustered across continents. The bottom panel shows the same data after spatial thinning with a 5° minimum distance, resulting in far fewer, more evenly spaced points that preserve broad geographic coverage while reducing clustering.
Results of the R function microbenchmark on three versions of a thinning algorithm: the naive algorithm written in plain R shows a median runtime of 49 seconds, a C plus plus version shows a median runtime of 11 seconds, and a C plus plus version with an optimized algorithm shows a median runtime of 0.1 second
Claude Code spinning five instances of an agent specialized in debugging and testing one R function at a time.
Snapshot of the kitty terminal showing two instances of Claude Code, one working on code optimization, and other working on a test suite.
Screenshot of an ECMWF forum announcement titled ‘Scientific and technical support for AgERA5 ending from 31-12-2025.’ A staff member named Michela states that scientific and technical support for the Agrometeorological indicators from 1979 to present derived from reanalysis (AgERA5) dataset will no longer be available from 01-01-2026 and is not recommended for production use. The post notes that users will be informed if support is reestablished and provides a link to the ECMWF Support portal for enquiries.
RStudio editor showing R code that builds a recipe with step_collinear(), defines a linear regression model using parsnip, constructs a workflow, fits it to vi_smol, and then calls broom::tidy() on extract_fit_engine(vi_workflow). Below the code, the Viewer displays a tibble with model coefficients
Screenshot showing R code calling collinear::collinear() with a data frame, response "vi_numeric", and numeric predictors. The printed output shows: response name, filtered data frame dimensions (610 rows, 12 cols), preference order with the ranking method (f_numeric_glm) and top predictors (growing_season_length, soil_ph, swi_mean, etc.), the final selection of 11 predictors after filtering, and ready-to-use model formulas for linear (lm) and smooth (GAM) models.
Scatterplot showing results from 10,000 simulations validating collinear's adaptive threshold system. X-axis shows input correlation (75th percentile), Y-axis shows output maximum VIF after filtering. Points are colored by number of input predictors (25-100). Two horizontal dashed lines mark VIF = 2.5 and VIF = 7.5. The plot shows that output VIF stays bounded within this range across diverse correlation structures, with higher input correlations producing higher but still bounded output VIF values.
Linkedin snapshot showing a data science lead job in Germany paying 14 dollars per hour. It's no longer accepting applicants, and shows that zero people clicked "apply" .
Results of devtools::test() showing the 799 successful unit tests of the R package {collinear}.
R function named wtf() that calls itself in a self-recursion pattern.
Barplot showing the monthly downloads of the R package collinear, with a monthly average of 384 downloads, and a clear increase after November 2024, resulting from release 2.0.
Bar plot showing the download history of the R package distantia from summer 2019. It shows a steady average of 480 downloads per month, with a slight decrease after the release of version 2.0.0 in January 2025.
A rubber duck wearing sunglasses and drinking a mojito or whatever.
Scatterplot showing the relationship between pairwise Pearson correlation and variance inflation factors across 10000 repetitions on randomized datasets. The data is organized as a sigmoidal curve, and colored according to the Jaccard similarity between the variables resulting from a multicollinearity filtering using a given correlation and a given variance inflation factor.
Snapshot of a Jenkins job running a Dropbox backup.
Snapshot of a workflow built with the defunct software Kepler.
Topographic map showing the location of several plant populations, and lines between them representing potential dispersal routes.
Three D map of eastern Andalusia showing the locations of several plant populations, and lines between them representing potential dispersal routes.
Topographic map showing the location of several plant populations, and lines between them representing potential dispersal routes.
Running testthat unit tests in parallel with Rstudio.
Snapshot of Rstudio Server running unit tests in parallel. The upper left panel shows an R function, the upper right panel shows htop on the console, the 'build' panel on the lower left shows the testing progress, and the lower right panel shows the file explorer.
Snapshot of my laptop's desktop. On top it shows the guake console running htop. On the lower right corner it shows a thunar window, while the lower left corner shows the desktop background, a still of the warship Rocinante (also the laptop's name) from The Expanse.
Screen capture saying: "Nobody gives a flying fuck about you. Not in the existential, "nobody loves me" sense. That would be melodramatic (and not true). What I mean is this: everyone else is too busy being the main character in their own lives to devote more than a passing thought to yours. And that's liberating."
Movie poster of the movie "Dumb and Dumber", with the actors Jim Carrey and Jeff Daniels.
Home battery showing 55% charge after 17 hours of blackout.
Home battery in use to deal with countrywide energy outage in Spain.
Forty years old cactus showing off eighteen beautiful pink flowers.
Picture of an inflatable stand-up paddle board (also known as iSUP) in a beach of the Alboran Sea.
Snapshot of a mind map generated by Google's notebookLM from a couple of links to a personal website and github pages.
A beach by the Alboran Sea in Southern Spain, full of trash returned back by the waves.
Private message saying: @schizanon.bsky.social has added you to the "does not get bitcoin/nft/ai" moderation list.
Blonde doggie using her bed as a caparace
Code snapshot of a variable importance analysis that explains the contribution of individual variables to the dissimilarity between pairs of time series. It is performed with the function momentum(), which produces a data frame with the columns x and y with the names of the compared time series, a column variable with the name of each variable, a column importance with the importance scores, which are negative when the variable contributes to similarity, and positive when the variable contributes to dissimilarity.
Multipanel diagram showing a dynamic time warping solution between two time series representing the flight paths of two albatrosses in the pacific over a period of five months. The time series have the variables longitude, latitude, speed, heading, and temperature. The time warping runs close to the diagonal, meaning that the two time series have relatively similar shapes.
Snapshot of R code showing the following code blocks: 1: the loading of the R packages distantia and DT; 2: the initialization, aggregation, and scaling of a time series object with flight paths of albatrosses in the Pacific using the functions tsl_initialize, tsl_aggregate, and tsl_transform; 3: dissimilarity analysis with dynamic time warping and restricted permutation tests performed with the function distantia; 3: usage of the function DT::datatable to show the resulting data frame, which contains the columns x and y with the names of the compared time series, the column psi with their respective dissimilarity scores, and the column p_value with score of the restricted permutation test.
Snapshot of R code showing the install of the R package distantia via the command install.packages(), and the version of the newly installed package as printed by the command packageVersion().
Snapshot of a failed attempt to submit version 2.0 of the R package distantia to CRAN.
Snapshot of completed unit tests of the R package distantia. It shows 251 tests passed in 29 seconds, and the text "Your tests deserve a gold medal".
Picture of the book Project Hail Mary, by Andy Weir.
Cloudy sunrise at the Alboran Sea.
Picture of the book "A Philosophy of Software Design", by John Ousterhout. There is also a duck with sunglasses in the picture. A rubber duck, not a real one. But we still talk from time to time. So it might be real to me anyway. Still, the important part is the book, not the duck, or the sunglasses.
Snapshot of the htop application showing 100% CPU usage in a ubuntu machine with sixteen cores, and a RAM usage of 15 out of 62 gigabytes.
Snapshot