Architecture, real systems, product decisions and technical lessons. Everything reviewed with my own judgment before it ships.
240 articles published
With a small, reproducible dataset in Jupyter, I compare what XGBoost actually does differently from scikit-learn's RandomForest — and why the answer isn't "which one is better" but "which problem is this for."
Sep 16 2026 · 8′ · Tutorials · machine learning · python
Netron lets you inspect the architecture of any ML model — no Jupyter, no code, no drama. ONNX, PyTorch, TensorFlow: open it and see everything.
Jul 05 2026 · 5′ · Experiments · tooling · machine learning
XGBoost isn't a trend — it's the algorithm that won hundreds of ML competitions on tabular data. Why it's still the mandatory reference in 2025 and when you should reach for it.
Jun 29 2026 · 5′ · Experiments · machine learning · open source
PyTorch showed up in 6 independent awesome lists and the reason is simple: it won. This isn't hype — it's infrastructure. Here's why it made our list and when it actually makes sense to use it.
Jun 26 2026 · 5′ · Experiments · machine learning · deep learning
TensorFlow isn't sexy in 2025, but it's still the serious infrastructure behind deployment at scale. Why it made the list and when you actually need it.
Jun 23 2026 · 5′ · Experiments · machine learning · deep learning
lode reimplements DVC's hot path in Go with a non-negotiable invariant: byte-identical compatibility with DVC 3.x. Static binary, parallel hashing, state DB that avoids re-hashing. No migration, no lock-in. But pipelines are out of scope and benchmarks have context. Here's why that perimeter is an honest technical decision, not a limitation to hide.
Jun 21 2026 · 7′ · Experiments · machine learning · open source
I followed the viral 241-point HN tutorial and documented every dollar spent, every GPU hour, and every disappointment. My thesis: training an LLM from scratch in 2025 is a valid technical exercise, but calling it "an alternative to Claude" is lying to yourself.
May 05 2026 · 9′ · Experiments · LLM · ia
The Python ML ecosystem has a structural problem that Node and Rust solved years ago: the transitive dependency chain of a single ML library can exceed 200 entries, most without verifiable cryptographic signatures. I simulated the same vector against my own stack — and what I found is not reassuring
May 01 2026 · 8′ · Experiments · devops · produccion
m2cgen (PyPI, Python) converts trained scikit-learn models into standalone Java, Go, C#, Rust and 8+ other languages — no Python runtime, no API, no serialized binary. Here's how it works and when to skip it.
Apr 18 2026 · 5′ · Experiments · machine learning · open source