I'm a geoscientist who solves problems with Python — from subsurface characterization to statistical detection of flawed research. My work sits at the intersection of domain expertise, data science, and AI-assisted reasoning. I build tools that make quantitative skepticism systematic: checking assumptions, questioning numbers, and catching the patterns that don't survive scrutiny.

Currently developing agentic AI applications for document extraction in geotechnical engineering.

Board member at Software Underground — open-source geoscience community.

Contact: matteo@mycarta.ca | Blog: mycarta | Twitter: @my_carta

A Python package for systematically screening published research for statistical red flags and methodological issues. Four-tier detection system, from quick API lookups to deep data analysis.

What it catches: p-value inconsistencies, impossible descriptive statistics (GRIMMER), spurious correlations in high-dimensional data, underpowered studies masquerading as significant findings, causal claims without evidence, and language escalation from "association" to "effect."

What makes it different: The package includes 12 structured skill files (~2,500 lines) that teach an LLM when to reach for which tool and how to interpret what it finds. The reasoning frameworks — Mill's Methods inverted as audit checks, Fermi sanity estimation, logical fallacy detection — are as important as the code.

>>> from bullshit_detector.spurious import P_spurious, r_crit
>>> P_spurious(r=0.50, n=21, k=100)
0.884   # 88.4% chance this correlation is spurious
>>> r_crit(n=21)
0.433   # minimum r to even consider at n=21

>>> from bullshit_detector.power import achieved_power
>>> achieved_power(effect_size=0.5, n_per_group=16)
{'power': 0.293, 'adequate': False}
# 29% power — less than a coin flip. If significant, probably a false positive.

9 modules, 104 tests, 7 example scripts, 12 skill files. Published on PyPI.

pip install bullshit-detector

Links: PyPI | GitHub | LinkedIn launch post

~10 years of research, community contribution, and accessible visualization

Poor colormap choices create false patterns in scientific data. I've spent a decade developing methods to evaluate colormaps perceptually, sharing the results through papers, blog posts, and interactive tools.

Try the interactive app — see how bad colormaps distort geophysical data.

Impact:

• Evaluation paper — Top 30 most-downloaded SEG papers (2010-2020)
• Cited in official Matplotlib documentation
• Blog posts rank #1 on Google for "perceptual rainbow" and "perceptual palette images"
• Knowledge exchanges with Peter Kovesi, Kristen Thyng, Bernice Rogowitz, Fabio Crameri
• January 2026: modernized to Panel 1.8.5 + Python 3.12, added cmcrameri and cmocean collections

Links: Live app | Source | Paper | SEG Wiki

Teaching LLMs quantitative reasoning through structured laws and worked examples

A framework of 17 Laws (mechanical + estimation) that teaches AI models to decompose problems, bound unknowns, proceed with imperfect information, and know when to ask for help. Developed over three years, tested on 11 problems with a 6-criteria scoring rubric. Codifies methodology from Weinstein's Guesstimation books into explicit rules for human-AI collaboration.

Connected to bullshit-detector as the "Fermi sanity" tier — order-of-magnitude plausibility checks on reported claims.

Blog series: Part 1: The Problem That Wouldn't Compute | Part 2: Permission to Guess

Anti-sycophancy guardrails and structured prompting for rigorous AI work

Working practices for getting honest, useful behaviour from language models — learned the hard way, across real projects where sycophancy, silent data corruption, and fabrication under questioning cost real time. Two blog posts document the failure modes and the systems I built around them.

The core principle: AI tools that agree with everything you say are worse than useless for analytical work. Configure them to challenge assumptions, not confirm them.

Blog: Operational Discipline for LLM Projects | Standing in the Middle of Intelligence | GitHub

The intellectual ancestor of bullshit-detector

A methodology for combining domain expertise with statistical rigour when evaluating published results. Don't just accept visual/qualitative claims — back them up with custom error flags, bootstrap confidence intervals, influence plots, and distance correlation. This detective approach evolved directly into the bullshit-detector package.

Links: Blog post | Notebook 1 | Notebook 2

When 19th-century philosophy and neural networks agree

A drilling problem from a CSEG talk: seven wells, five seismic attributes, four with mud loss problems. Three approaches converged — Mill's Methods of Induction (1843), a simple neural network, and domain expert Lee Hunt — all pointing to the same attributes. The pragmatic insight: the goal isn't finding "the cause," it's building a defensible decision rule under uncertainty.

Links: Blog post | Notebook

How does a nearly blind robot cover every cell in a room? Picobot can only sense its four immediate neighbours — no map, no memory. In 2015, working through Harvey Mudd's CS materials on my own, I optimized the empty room solution from 7 to 6 rules and the maze from 16 to 12. The key insight: the X move (stay put) enables state transitions that let you reuse rules instead of duplicating logic.

Revisited in 2025 with a Python simulator, exhaustive verification, and proper documentation.

Links: Blog post | GitHub

A free, offline, no-ads mobile game based on Math Games with Bad Drawings by Ben Orlin. Plant flowers, spread seeds, try to cover every square before the Wind leaves you with gaps. Built with the author's knowledge and encouragement.

Single player vs computer Wind (Easy/Hard) • Interactive tutorial • Installable as offline PWA • No accounts, no tracking

Includes a JavaScript tutorial for Python developers walking through the game engine.

Play it — works on phone, tablet, or desktop.

Links: Play | GitHub

How to evaluate and compare color maps — The Leading Edge 34(8), 2015. Top 30 most-downloaded SEG papers (2010-2020).

Mapping and validating lineaments — The Leading Edge 34(8), 2015.

Introduction to Classification with SVMs — The Recorder (CSEG), expert-reviewed.

Keep on improving your geocomputing projects — Chapter in 52 Things You Should Know About Geocomputing, 2019.

Software Underground — Board Member (2024-present)

mycarta blog — 65+ technical posts on geoscience, visualization, data science, and AI.

Conference talks: Transform 2020 (colormap tool), TRANSFORM 2021 (FRIDA hackathon)

Stack Exchange: Land percentage in Northern Hemisphere — combining map projections with Python to answer a deceptively simple question.

Last updated: March 2026