Everything we have published, newest first.
Wrap the model call, count tokens per step, and turn one run into a cost estimate you can take to leadership.
Sep 29, 2026
•
8 min read
Build a majority-class and a mean baseline in scikit-learn, then use them to decide if the real model is worth shipping.
Sep 27, 2026
7 min read
Two fraud models tie at 99.5% accuracy, and only one is useful. Ask what each mistake costs to pick the metric.
Sep 25, 2026
6 min read
A fine-tuned returns bot keeps quoting the old 30-day window; here is a rule for what belongs in weights versus prompts.
Sep 22, 2026
Type hints don't check LLM output at runtime. Here's how one Pydantic model catches a bad ticket at the door instead of three files later.
Sep 20, 2026
Clear instructions, few-shot examples, chain-of-thought and role prompting: what each one fixes, and how to pick the right one for the task.
Sep 18, 2026
3 min read
Why the deprecated RetrievalQA chain hides too much, and how piping small LCEL pieces gives you streaming, parallel retrieval, and debuggability.
Sep 16, 2026
Turn fine-tuning from a manual research event into a CI/CD pipeline, and every customer correction becomes a moat that a bigger generalist model can't copy.
Sep 4, 2026
