Publications
Recherches sur les agents IA, backtests et audits de prévisions de marché, et architecture de la plateforme.
IA et entreprise
Ten clients at $8,000 a month is $960,000 a year. The real profit arithmetic, six categories already working with the companies doing them, three pricing models, and a step-by-step from zero to first retainer.
10 min de lecture9 sections
IA et entreprise
An AI agent runs a loop - observe, decide, act - and a chatbot does not. What that difference demands in scope, records and stop controls, where agents genuinely win, and four questions that evaluate one faster than any demo.
5 min de lecture6 sections
Pièges des modèles
Regime shift is the most common way machine learning fails in production, and backtests average it away. A worked example with real numbers from our own board, and four habits that catch it before the results do.
5 min de lecture6 sections
IA et entreprise
For most people AI changes the cost side of a household long before the income side. What genuinely moves - fees, comparison, gated expertise - what it puts at risk, and why "let AI trade for you" is the weakest version of the claim.
5 min de lecture6 sections
IA et entreprise
AI removes steps for tasks where being wrong is cheap and visible, and adds a verification tax where being wrong is expensive and invisible. A single question sorts almost every case correctly.
5 min de lecture7 sections
Crypto et marchés
MEV is one of the few places where machine learning has a defensible job, because the task is simulation and classification rather than price prediction. But private order flow makes every public-data study a lower bound of unknown tightness.
6 min de lecture6 sections
Crypto et marchés
Most depegs that matter are plumbing wearing the costume of solvency. What machine learning can genuinely do with stablecoin data, why depeg prediction breaks on rare events, and how supply works as a regime descriptor rather than a signal.
6 min de lecture6 sections
Crypto et marchés
Funding rates and open interest describe how levered a market is and what that position costs to hold. They forecast the size of coming moves, not the direction - and the data is full of traps that quietly wreck studies. What AI can honestly extract.
6 min de lecture6 sections
Crypto et marchés
On-chain data is the only complete dataset in markets - no survivorship bias, no vendor filter. It is also built on unvalidated address labels and a handful of independent cycles. What AI can honestly do with a blockchain, and what it cannot.
7 min de lecture7 sections
Évaluation des prévisions
The instinct to grade a forecast by whether it came true is the wrong test. Calibration - when a model says 70% it happens about 70% of the time - is the property that separates a real forecaster from a confident guesser, and it is measurable.
7 min de lecture7 sections
Crypto et marchés
Direction is the AI forecast everyone wants and a liquid market defeats. How far an asset can move is the one machine learning can make - and prove. Why sigma-root-t is wrong in both directions, and what coverage really tests.
7 min de lecture7 sections
Crypto et marchés
Everyone chases AI that predicts price. The place machine learning genuinely earns its keep is risk management — volatility, position sizing, drawdown, liquidation. Measurable, unglamorous, honest.
9 min de lecture7 sections
