Applied AI engineering

AI integration services are applied AI engineering: I add large language model features, retrieval over your documents (RAG) and agents to real products, together with the evaluations, guardrails and cost controls that keep them useful in production. It is for product teams in the US, EU, UK, Japan and Korea who want an AI feature their customers will rely on, not a demo.

The engineering is in everything around the model: what it reads, what it may do, how its answers are checked, and what each request costs.

It suits teams with large document collections, support histories or knowledge bases; teams whose prototype worked in a demo and failed on real inputs; operations teams with repetitive work across several tools; and SaaS companies whose customers want to use the product from an AI assistant.

It does not suit features where a wrong answer is dangerous and no person can review it, fully autonomous agents with unrestricted access to production systems, or work with no clear definition of a correct result. If that definition doesn't exist yet, writing it is the first job.

Four kinds of AI work come up most, and each has its own page.

A RAG feature answers questions from your own documents and cites them. It is the usual first step, with a one-week scoping sprint: AI features in your product.

An agent acts inside your tools with scoped permissions, an audit log and a person confirming risky actions, and an MCP server lets your customers' assistants use your product the same way: AI agents and MCP servers.

An automation moves data between your tools on a schedule or a trigger, with AI steps where they help and a person approving where they should: workflow automation and internal tools.

Coding agents write and change code in your own repository, and the work is setting them up to ship safely: AI coding agents for engineering teams. Once any of these is live, LLM evals, guardrails and cost control measures whether it is getting better or worse.

"It answers well in the demo and makes things up with real customers." "We changed the prompt and nobody knows whether it got better." "The bill grows faster than usage." "Security asked what happens if a document tells the model to ignore its instructions, and we didn't have an answer."

Each of these is a measurement problem first. Without real questions and agreed answers, every change is a guess.

You get the feature integrated into your product, an evaluation set built from questions your users actually ask, scored on every change, guardrails against prompt injection and data leakage, and cost controls such as caching, model routing and usage limits. Answers are grounded in retrieved sources and cite them.

AI features in your product covers retrieval, extraction, classification and drafting. AI agents and MCP servers covers agents that act inside your tools with scoped permissions and an audit log, and Model Context Protocol servers that make a SaaS usable from AI assistants. LLM evals, guardrails and cost control is for a feature already in production. The workflow-automation and AI-coding offers below extend the same approach to internal operations and to engineering teams.

With a one-week, fixed-price scoping sprint on your real data. I build a working spike, run it against questions your users actually ask, and report measured quality, cost per request and the risks, before anyone commits to a build.

The same discipline ran the Zyrobe pipeline, where cost per image was a design constraint from the first day and an upload became a finished product image in 3–8 s on average.

For EU teams, two things shape the design. Personal data sent to a model provider is a transfer like any other under the GDPR, so where the data goes and under which terms is settled first, and open-weight models can run inside your own infrastructure when it must not leave. And Article 50 of the EU AI Act requires people to be told when they are interacting with an AI system, unless that is obvious, and AI-generated content to be marked in a machine-readable way; those duties have applied since 2 August 2026, with a transition to 2 December 2026 for marking by generative systems already on the market before that date. For South Korea and Japan, the work is usually AI features shipped with documentation of how they are evaluated, what they cost and what happens when they are wrong. Which rules apply to your product is for your counsel to confirm.