Practice
LLM application development and AI software
Language-model features built into business software, with the evaluation set written before the feature. This is active development rather than a mature product line, and it is described that way.
An AI feature demos well and then becomes impossible to change. A prompt is edited, a cheaper model is swapped in, a rule moves, and nobody can say whether the result improved. That is a testing problem before it is a modelling one, so NextWR starts with the evaluation set: real inputs with the answer that would be accepted for each.
The work that follows is ordinary software engineering. Generation workflows that produce structured output rather than prose, so the result can be stored, compared and audited. Evaluation runs wired into continuous integration so a change reports its effect. Guardrails on what the model is allowed to decide, and a human step wherever the output carries consequence.
Model and provider choice follows cost, privacy and latency rather than fashion, and it is a decision taken per project rather than once for the company. Where data cannot leave the client's infrastructure, open-weight models running locally are the starting point, with the quality trade-off measured on the same evaluation set rather than assumed.
Local and edge inference is an area of active work rather than a finished offering. It is described here because it is genuinely being developed, and it will be described as mature only when it is.
What is included
- An evaluation set built from real cases, before the feature is built
- Generation workflows producing structured output rather than prose
- Automated evaluation wired into continuous integration
- Guardrails, and a human review step where output carries consequence
- Model and provider selection by cost, privacy and latency
- A maintenance arrangement, because retrieval quality drifts as data changes
Not included
- Training a foundation model from scratch
- Buying or scraping training data on a client's behalf
- Accuracy claims that cannot be measured on the client's evaluation set
Common questions
How do you know the output is right?
Nobody does until there is an evaluation set. We build one from real cases, run it on every change and report the result. Without it the feature is a demonstration rather than something that can be maintained.
Can it run without sending data to a third party?
Yes, with open-weight models on infrastructure the client owns. It takes longer to set up and is easier to defend in a privacy review. We measure what it costs in quality for the specific task before the choice is made.
Tell us what you are building.
Write in English, French, Romanian or Arabic. Every message gets a written answer within one business day.