Researcher · Brasília, Brazil

Pedro Carvalho Brom

Researcher, developer and consultant in Computer Science, with a background in Statistics and Applied Mathematics. Currently a PhD candidate in Informatics, specializing in generative artificial intelligence, working with advanced analytical methods in R and Python to build data-driven solutions. The work bridges statistical modeling, Artificial Intelligence and business technology, with a focus on AI solutions that are mathematically sound, computationally efficient and viable in production.

LinkedIn articles and posts

Long-form articles and notable short posts published on LinkedIn.

See LinkedIn profile

Featured work

Packages in production and research prototypes. Papers are listed under Publications. Filter by category.

Open source

gpumetropolis

R package on R-universe: a Metropolis-Hastings sampler with a vendor-agnostic GPU-portable log-density kernel (Rust and CubeCL); pre-registered factorial benchmark against five mainstream MCMC packages.

View on R-universe

Open source

autoresearch

Autonomous research loop with a local LLM critic; a generalization of karpathy/autoresearch.

View on GitHub

Open source

community-skills

Community hub for wrapping libraries as agent-callable skills, in R, Python and Julia.

View on GitHub

Open source

NLPMetricLab

Multilingual NLP metric toolkit (BLEU, CHRF, TER, semantic similarity); companion app to the FITEE back-translation study.

View on GitHub

Open source

trace-llm

Code, data and prompts for the TraCE-LLM study on traceability in LLM pipelines.

View on GitHub

Open source

2-Steps-RAG

Two-step RAG for metadata filtering and statistical LLM evaluation; research code.

View on GitHub

Open source

SimpleChatGPT

Simple ChatGPT integration for the GNU/Linux shell.

View on GitHub

Open source

leidebenford

Application of Benford's law to dictionary data, in R.

View on GitHub

Open source

bgev

R package for the bimodal generalized extreme value model.

View on GitHub

Open source

bgumbel

R package on CRAN for the bimodal Gumbel distribution, a three-parameter model for extreme data with one or two modes.

View on CRAN

Open source

amazeR

Maze generation and random-walk simulation in R.

View on GitHub

Open source

olxbuscador

R scraper for collecting listing data from the OLX marketplace.

View on GitHub

SaaS

SOTA

AI-assisted platform for the academic research workflow: manuscript review, journal selection, reference-integrity checking, and research discovery through author metrics, semantic search and knowledge graphs.

Coming soon, July 2026

Publications

Journal and conference papers.

See all on Google Scholar

Software and datasets

Citable research software and data, each with its own DOI. Most recent release of each project.

See all on ORCID

All repositories

Every public repository, most recent first, including course material, experiments and small tools.

See all on GitHub

harnessCurated agentic harnesses for R professional roles; launches CLI coding agents in a terminal tab with role-specific skills, prompt and audit-first manual execution
pcbrom.r-universe.devR-universe of Pedro Carvalho Brom
immba
ci_stats
MMBN-CAS
perceptron-mlp-cnnPerceptron, MLP and CNN implementations.
stat_courseLecture notes for an Applied Statistics course.
bt-conference
webist_cep
aulasrose
maria
pds
fingerprint
LGA2011List of processors compatible with the LGA2011 socket.
q2b-l2
balduRsgateAutomates rolling for character attributes in Baldur's Gate II: Enhanced Edition.
energy
cleandRoid
facebotRemoveOldInv
internet-back-beep
atalhos
keymapChange
speedmonitoR
facebook.bot.curtir.paginaBot to invite Facebook friends to like a page.
whatsapp_botWhatsApp bot.
RSeleniumDriverA working fix for the RSelenium driver loading problem.
RwhatsappWhatsApp bot written in R.
facebook.bot.parabensBirthday message bot.
desafios_RDojoSolutions to the RDojo.com.br challenges.
icc_est_r
RProgIntroduction to Programming with R course.
CPSKCapstone project, SwiftKey.
CourseraFPDataProducts
Getting-and-Cleaning-Data-Course-ProjectGetting and Cleaning Data course project.
CourseraRegressionModelsCoursera Regression Models course project.
CourseraMachineLearning
IPRCourse material: Introduction to Programming with R, IFB.
Estatistica_ComputacionalSimulation material for a Computational Statistics course.
Experimento_FatorialFactorial experimentation and optimization workshop, IFB JEPE 2016.
datasciencecoursera

Blog

Technical notes on statistics, AI and R.

See all posts

13 JUL 2026 · LLM

The model copies the number correctly and then concludes wrong: move the reasoning into code

A 2B model given a raw record copies the numbers faithfully and then draws the wrong conclusion from them: at the 4th percentile it still says intermediate. So the pipeline moves the reasoning into deterministic code and leaves the model only the prose, with numbers as symbols the harness substitutes. The paired ablation puts a number on it: remove the mediation and 30% of readings carry a factual defect, against 0% with it, McNemar p = 2.9e-11. What breaks in a small model is not transcription. It is ranking.

Read the post →

FAQ

Quick answers about projects, papers and how to reach me.

Who is Pedro Carvalho Brom?

Researcher, developer and consultant in Computer Science with a background in Statistics and Applied Mathematics. PhD candidate in Informatics, specializing in generative artificial intelligence. Based in Brasilia, Brazil.

What is autoresearch?

An autonomous research loop with a local LLM critic: the agent generates, critiques and refines its own outputs across iterations, with a model-side judge instead of a human-in-the-loop on every cycle. Source code at https://github.com/pcbrom/autoresearch.

What is community-skills?

A community hub for wrapping libraries as agent-callable skills, with implementations in R, Python and Julia. The goal is to give LLM agents typed, deterministic interfaces to existing scientific tooling instead of letting them generate fragile code on the fly. Source code at https://github.com/pcbrom/community-skills.

What is TraCE-LLM?

A published methodology for evaluating trait consistency in large language models: behavioral stability across prompts and the adversarial compensation effect. Brom, Oliveira and Weigang, Journal of the Brazilian Computer Society 32:1 (2026), DOI 10.5753/jbcs.2026.6343.

What is gpumetropolis?

An R package distributed via R-universe: a Metropolis-Hastings MCMC sampler with a vendor-agnostic GPU-portable log-density kernel implemented in Rust with CubeCL. Includes a pre-registered factorial benchmark against five mainstream MCMC packages. Available at https://pcbrom.r-universe.dev/gpumetropolis.

What is bgumbel?

An R package on CRAN implementing the bimodal Gumbel distribution, a three-parameter model for extreme-value data with one or two modes. Companion to the 2023 article in the Austrian Journal of Statistics. Available at https://cran.r-project.org/package=bgumbel.

What is SOTA?

An AI-assisted platform for the academic research workflow: manuscript review, journal selection, reference-integrity checking, and research discovery through author metrics, semantic search and knowledge graphs. Launching July 2026 at https://sota.pcbrom.com.