AI research notes with open numbers: methodological articles we publish because we can defend them, each with a documented truth-check, clear sources and a companion repository.
Recalculated,
not retold: AI Research Notes
The louder the AI promises, the rarer anyone shows the maths behind them. We look behind the curtain and do the maths.
-
01 Research Methodology
A Truth-Check Protocol for AI research output
How we publish nothing at myBytes we cannot defend. Seven steps from AI claim to publication, a tier system for sources and a truth-check on our own claims.
Guido Winger·11 min read· -
02 EUDR Supply-Chain Intelligence
EUDR without gut feeling
How an auditable per-pixel risk mask works. Hansen × FDP plantation, two example regions, a per-pixel audit trail.
Guido Winger·12 min read· -
03 EUDR Supply-Chain Intelligence
Six months before the new EUDR deadline
What really changed since December 2025 and what did not. What satellite monitoring has shown since, with a roadmap across four quarters.
Guido Winger·11 min read· -
04 Soft-Commodity Volatility
The single-GARCH limit on soft commodities
A GJR-GARCH-t passes the VaR discipline on four ICE soft commodities but delivers no early warning. What follows from that, with a reproducible companion repository.
Guido Winger·8 min read· -
05 Research Methodology
Seven questions decide whether your AI project fails
A reproducible data-quality audit for mid-cap data teams: seven dimensions, a traffic-light verdict, an open-source tool. Use only on backups or samples.
Guido Winger·9 min read· -
06 Soft-Commodity Volatility
The second layer: what a regime model really delivers in lead time
A hidden Markov model on the same four soft commodities, at the identical pre-specified endpoint as the GARCH baseline. Which lead-time figures you may believe, and why.
Guido Winger·10 min read· -
07 EUDR Supply-Chain Intelligence
EUDR compliance does not need prettier maps. It needs defensible pixels.
Why satellite maps alone do not create EUDR compliance, and why what counts is traceable, reproducible decisions at the pixel level.
Guido Winger·9 min read· -
08 Generative AI
Why 95% of GenAI pilots fail
Seven methodological causes why GenAI pilots fail, and seven concrete countermeasures for CFOs, COOs and CDOs, all transferred from the replication-crisis diagnosis.
Guido Winger·10 min read· -
09 Fashion Intelligence
What fashion returns really cost
Recomputed on 2.33M real order line items: returns prevention in fashion sits on the economic edge. An honest cost range, the break-even and the real size lever instead of a millions promise.
Guido Winger·8 min read· -
10 Fashion Intelligence
What size recommendation actually does against returns
The 25% claim of personalised size recommendation, honestly measured: hit rate 28.8 vs 19.1%, realistically around 2% of all returns, theoretical ceiling 6.9%.
Guido Winger·6 min read· -
11 E-Commerce Intelligence
Why near-perfect conversion models mostly retell the basket
Reproduced on real shop data: the near-perfect conversion AUC is mostly basket tautology. What is actionable is the honest early prediction at 0.85, not the 0.96 at session end.
Guido Winger·6 min read· -
12 Sovereign AI
Sovereign AI, honestly priced: what on-premises really costs
Both sides of the ledger completed: the break-even versus the cloud sits at around 574 million tokens per month, and the sovereignty premium costs about EUR 4,700 a month. A model calculation from 110 dated price sources.
Guido Winger·10 min read· -
13 Computer Vision
Malaria detection with deep learning
A compact CNN detects malaria in blood-cell images at 99.77% sensitivity: only 3 missed infections in 1,300. Eleven model variants compared, with a reproducible companion repository.
Guido Winger·12 min read· -
14 Customer Analytics
Customer segmentation with clustering
Five clustering algorithms compared, K-Means (k=3) wins: three actionable customer segments from 2,240 customers, with a reproducible companion repository.
Mariusz Pianowski·10 min read· -
15 E-Commerce Intelligence
Amazon product recommendations: which method wins when
Four recommender methods compared on 65,290 Amazon ratings. Tuned SVD wins on RMSE (0.8822), rank-based on the cold start, so the best model depends on the scenario. With a reproducible companion repository.
Guido Winger·9 min read· -
16 Customer Analytics
House price regression: what an R² of 0.77 really tells you
Linear regression built with discipline: log transformation, VIF against multicollinearity, four assumption checks, 10-fold cross-validation. Why the 5% MAPE is not a 5% price error.
Mariusz Pianowski·9 min read· -
17 Research Methodology
Two analysts, one dataset: where the numbers agree and the recommendations diverge
Two independent analyses of 1,898 orders: identical numbers, opposing recommendations. Why reproducibility is the entry ticket and the recommendation is the actual modelling step.
Guido Winger & Mariusz Pianowski·10 min read· -
18 Generative AI
AI slop: why the web is filling up with junk
Where the term comes from, documented examples from Clarkesworld to Facebook, and the patterns that help you spot mass-generated content.
Guido Winger·7 min read· -
19 Generative AI
AI content labeling: who must comply with EU AI Act Article 50, and which exemptions apply
Article 50 for decision-makers: the four sets of obligations, the exemptions, the deadlines and an implementation that survives everyday operations.
Guido Winger·8 min read· -
20 Generative AI
A complete web application from AI: the self-experiment
An AI environment builds a browser groovebox with its own drum model, and the logbook shows what the AI did well and where it failed.
Guido Winger·10 min read·
No articles for this selection.