AI and agent systems

An AI-based project starts with the question of what needs to improve and how to measure it, not with choosing a model. Only afterwards does the work move down to integration, prompts and how results get evaluated.

Integrating a language model means working with its API, designing the context, handling errors and controlling the cost of requests. The model's output gets treated as input to be checked, not as a finished answer.

Agent tooling gets built around the Model Context Protocol. An agent gets a narrow scope, explicit permissions and a gate that halts the work when the result fails a condition - the same pattern that guards the contract in the CI pipelines of other projects.

Automation starts with measurement, not with a tool: how many times a month someone performs a given task, how long one run takes and what happens when the data is unusual. The cheapest automation is often a plain script run on a schedule and wired into an existing API - a robot clicking through an interface is left as a last resort, because it breaks with every screen change. A language model steps in where the data is unstructured: classifying tickets, extracting fields from documents, drafting a first response. The model answers, a person decides - especially when the cost of a mistake is high, and a failed run lands in a queue for manual review instead of disappearing without a trace.

Data lands in one place from APIs, databases and web pages through scheduled jobs, and a test guards data correctness, not good intentions - a missing field or a duplicate stops the run instead of quietly entering a report. Traffic analytics run on Google Analytics 4 and Search Console, with consent configured the same way it is configured on this site.

Low-code works well for a prototype and a simple internal process, but the cost grows with the number of exceptions and the number of licences, and the business logic ends up locked inside a tool it is hard to move away from. Which process can stay no-code and which needs a dedicated application is judged case by case, and the exit from a tool that stops being enough gets managed without interrupting a team's work.

Out of scope: training proprietary models, computer vision and maintaining MLOps platforms - that is work for a team with its own infrastructure and training data.