Service · AI integration
AI integration services for products that already exist
Most AI features work beautifully in a demo and fall apart in front of users. I build the other kind: retrieval that returns the right thing, agents that call the right tool, costs that hold at volume, and a defined answer for what happens when the model is wrong.
I am a developer first. Most people selling AI integration cannot ship the application it has to live inside, and most engineers who can have never taken an LLM feature to production. This is both, on your existing codebase.
Book a call20 minWhat I build
The shapes this usually takes
RAG over your own data
Retrieval-augmented generation across your documents, tickets, or catalogue, with citations so answers can be checked rather than trusted. Chunking, embeddings, re-ranking, and the retrieval quality work that decides whether it is useful or noise.
Agents and tool calling
Models that call your APIs, read your database, and take actions inside your product. Tool design is most of the job here, because a badly described tool is the reason an agent picks the wrong one.
MCP servers
Model Context Protocol servers that expose your internal systems to Claude, Cursor, and other assistants, with authentication and scoping that stop them doing more than they should.
Classification and extraction
Intake that routes itself, documents that come back as structured data, and the unglamorous LLM work that removes the most manual hours.
Production
The demo is not the work
Getting a language model to answer once takes an afternoon. Everything below is what separates that from something you can put in front of paying customers, and it is where an AI integration project actually spends its time.
Evaluation before it ships
A test set and a measured pass rate, so a prompt change can be shown to help rather than assumed to. Without this you are shipping vibes.
Cost controls
Prompt caching, model routing, and per-user limits, decided against your real traffic. The difference between a $200 month and a $6,000 one is architecture, not usage.
Failure handling
Fallback models, timeouts, retries, and a defined answer for what the product does when the model is wrong, slow, or down. It will be all three eventually.
Abuse and rate limiting
Anything that puts a model behind a public endpoint is a bill somebody else can run up. That gets handled before launch, not after the invoice.
What it costs
AI features on an existing product start at $6,000, fixed in writing before any work begins. Smaller pieces run by the hour. Every engagement type and its starting price is published here.
Proof
Shipped, not prototyped
Growth Saloon
The marketing-and-software studio I build at day to day. We pair done-for-you marketing with custom platforms for the agency and the businesses it works with.
Play4TK
March Madness charity bracket pool honoring coach Tom Kowalski.
MintLooks
Flagship site for MintLooks, his web design + AI automation studio.
FAQ
Common questions
What does AI integration actually mean?
Wiring a language model into software you already run so it does a specific job: answering from your own documents, routing incoming work, extracting structured data, or taking actions through your APIs. It is engineering work on your existing stack, not a separate AI product bolted alongside it.
How is this different from adding a chatbot?
A chat widget is a front end. The work here is everything behind it: what the model can see, how it retrieves the right context, what it is allowed to do, what happens when it is wrong, and what it costs per user at real volume. A demo takes a couple of days. Production takes weeks, and the gap between them is where most AI features die.
Do I need a vector database?
Usually not a dedicated one. Postgres with pgvector, or Supabase, handles retrieval comfortably well past the point most products reach, and it keeps your data in one place. A separate vector store is worth it at scale or with unusual retrieval needs, not by default.
Which models do you work with?
OpenAI and Anthropic mostly, and the routing between them. Model choice matters less than people expect. Retrieval quality, prompt structure, and evaluation decide whether a feature works; swapping models rarely rescues a design that is wrong underneath.
What does it cost?
AI features on an existing product start at $6,000, and the full price is fixed in writing before any work starts. Smaller pieces run at an hourly rate. Every engagement type and its starting price is published on the web development page.
Can you work on a codebase you did not write?
Yes, and most of this work is exactly that. You have a product that already works and want an AI feature inside it that does not destabilise what is there.
Got an AI feature in mind?
A short call to talk through what you're building and whether I'm the right fit. If it isn't, I'll say so and point you somewhere better.