Engineering, AI and software delivery
What we have learned applying artificial intelligence in real development teams, with the sources on the table.
AI in development: what the data actually shows
One study measured 55.8% faster. Another measured 19% slower. Both are right, and the difference between them is the part that matters.
Read the articleAI code review without lowering the quality bar
People using AI assistants write less secure code and feel more confident it is secure. The gate needs to get stricter, not looser.
Read the articleRAG for a knowledge base: what survives production
Filling the context window can make the answer worse even with the right passage inside it. Position matters more than volume.
Read the articleEvals in CI: testing what is not deterministic
Equality assertions fail when the right answer can be written ten ways. Change the question the test asks and the suite becomes useful again.
Read the articleAI agents in delivery: where they pay off and where they do not
If you can draw the flowchart in advance, it is a workflow. Agents only when the path depends on what gets discovered along the way.
Read the articleSprint planning and estimation in the age of AI
The average may improve, but the range gets so wide the number stops being useful for a commitment. Task type starts mattering more than size.
Read the articleThe new technical debt: code you did not write
The old debt had someone who knew why they took the shortcut. The new one grows silently, and you discover it during an incident.
Read the articleAI-generated tests: coverage is not confidence
Coverage answers whether the line ran, not whether anyone checked the result. The right question is whether any test fails when you break the behaviour.
Read the articleFrom prototype to production: an LLMOps pipeline
If the prompt lives outside version control, you cannot reproduce an incident. And cost per interaction is a requirement, not a month-end report.
Read the articleObservability for AI features: latency, cost and quality
There is no counter for a "good answer". Quality is observed by approximation, and a dashboard without that axis gives a sense of control the system does not have.
Read the articleModernising legacy systems with AI as an archaeologist
That odd condition in the shipping calculation handles a real case nobody documented. Finding that out is the dominant cost of the project.
Read the articleA prompt is code: version it, review it, test it
An instruction is probabilistic guidance. Code is a guarantee. Confusing the two is the most expensive mistake in an LLM application.
Read the articleThere is no AI without data readiness
Data problems are barely visible early and compound later. Fixing at the source looks like waste and is the only thing that prevents the large rework.
Read the articleAI governance in Brazil: LGPD, Bill 2338 and the AI Act
A reviewer who approves forty cases an hour and never rejects anything is not a control, it is a stamp. A human in the loop needs time, information and authority.
Read the articleLLM application security: injection and tool abuse
The question is not where user input comes in, it is what the model can do if it believes the wrong thing. The answer is the size of the possible damage.
Read the article