Articles
Thoughts on how to build effective solutions, using hard-earned experience.
One conversation, two engines
Teams are trying to build entire journeys inside a chat model, and hitting the thing a model cannot give them: the same result every time. The answer is not to retreat to forms. It is one conversation with two engines behind it, where the model narrows and translates, a deterministic process executes and guarantees, and the customer never notices the seam.
The requirements nobody writes down
A good set of requirements now turns into a working system in days. That makes the half almost nobody writes down, how it should behave rather than what it should do, the thing that decides whether you have built something you can run. Usability, accessibility, performance and scale are architectural, and none of them can be added afterwards at a sensible price.
The router is the architecture
Classifying intent first and handing off to role-specific agents beats one engine trying to serve every query. But not because specialists are cleverer: because a good answer means something different for each intent, and because you can only evaluate what you have partitioned. The risk is that the router quietly becomes the monolith you were avoiding.
Your threshold is a function of the question
Retrieval confidence gates get calibrated against test questions written by people who already know the answer. Real customers ask worse questions, the same correct document scores lower, and the gate throws it away. The number was never the problem: an absolute cutoff was never comparable across questions in the first place.
The whisper game
Ask a coding agent to build something with several steps and you get a chain of small functions, each handing its output to the next. By the end of the chain the data is a rumour of what arrived. The fix is not a better prompt; it is orchestration, and one canonical entity persisted through the whole journey.
One truth, delivered to the edge
A single customer view that lives in a warehouse is a reporting asset. The same view, present at every touchpoint and every decision point, is a different product entirely. And once machine learning is in the path, context is the largest single lever on the quality of the answer.
Circles, not lines
You can engineer an aircraft that flies and lands, and still build the wrong aircraft. The gap between a working set of functions and a service worth running is a control loop: a measure that means something, and the authority to change the design when it moves.