Machine learning, NLP, and applied AI in production
A tutorial in code: how I stopped juggling six pools of French political text in three formats and built one queryable library — one folder per document, a Pydantic schema, a Parquet index rebuilt by a script, no database — and how to query it and define collections once it exists.
I stress-tested my newest, riskiest metric for weeks and shipped none for an old, boring one — a moral-foundations word list. A single mistranslated word had been inflating one whole moral dimension by about 2× for every speaker. The bug was not in the code. It was in the dictionary. A ported word list is code — test it like code.
Natural Language Inference trains the sentence embeddings everyone uses. It also works as a classifier that needs zero labeled examples. Here is the architecture, its history, and three very different results when I put it into production.
A practical tour of the z-score through two very different ML use cases — temporal anomaly detection on entity counts, and Monroe et al.'s log-odds-with-Dirichlet-prior for discovering discriminative vocabulary between two corpora.
AFK indexes journalism — so when Claude Sonnet silently rewrote verbatim quotes returned by tools, trust broke. Here's how I caught it with 6-gram overlap and fixed it with a step_callbacks accumulator on top of smolagents.
SQL sampling strategies, LLM pre-classification, a custom Streamlit review tool, and publishing to Hugging Face — the practical workflow behind building a training dataset.
Why your classification taxonomy matters more than your training data size. A practical guide through trial, error, and embedding analysis.
How I discovered SetFit, the framework that lets you build production-grade text classifiers with just a handful of labeled examples per category.
How I built a hyperparameter tuning system for BERTopic that can test hundreds of parameter combinations in minutes
How to use Claude Sonnet for NLP clustering ground truth. Use AI labeling instead of manual annotation.
A comprehensive guide to evaluating supervised clustering algorithms, bridging the gap from binary classification fundamentals to advanced clustering metrics through a practical, exploratory approach.
Automating a social media pipeline with n8n, enhance photos using Cloudinary's AI auto-cropping, generate poetic descriptions with Mistral AI, and handle errors gracefully.
Discover how I began tackling deep learning with a personal selection of must-read books—from Chollet’s deep dive to Glassner’s insights — alongside hands-on PyTorch tutorials, AI Engineering tips, and curated videos.
Step-by-step guide to building automated social media workflows with n8n. Features AI-powered image classification, API integrations, and smart content pipeline creation for effortless posting.
Step-by-step guide to configuring your NVIDIA Jetson AGX Orin with JetPack 6.2, installing TensorFlow GPU support, optimizing SSD usage, and setting up a Conda-based Python environment.
Step-by-step guide to running Pixtral-12B on a Jetson Orin AGX with JetPack 6.2. Covers cuSPARSELt setup, PyTorch and Transformers installation, and a working multimodal inference example on edge ARM hardware.
