Everyone says you need AI. Nobody has said where.
The board wants a plan. But nobody has said which of your tasks AI would improve, which it would make worse, and what each would cost per month.
We tell you where AI is worth it. Then we build it and keep it running.
For companies that want AI inside the software they already use, and an honest answer first about where it will pay off. We build assistants, search over your own files, multi-step AI tasks (agent pipelines) and MCP servers. Each one is connected to your real systems and tested on your real cases.
First step
System Review · fixed price · within a week
Often the honest answer is a simple rule or a basic link between two systems, and the report says so.
How every AI engagement starts
2018
Building and running live software since
1 week
System Review, delivered within a week
7
Products of our own we run in production
3
AI providers we can switch between
Where this usually starts
The pilot worked on the demo data.
On real questions it guessed, or showed someone a private document. The cause is usually missing access rules, stale data and no test questions, rarely the AI model.
What walks in
What people usually tell us about AI, and what we do about each
01 / 05
01
“My team spends hours finding things in our own files.”
Search that respects who may see what, points to the exact passage and says when there is no answer. A person checks any dates or clauses it pulls out.
Seen in: Law, Tax & Accounting Firms, Manufacturing & Export
02
“Customers ask the same questions every evening, or search and find nothing.”
We write the answers down as real pages and catalogue data, then let an assistant answer only from them and hand everything else to a person.
Seen in: E-Commerce & Retail, Restaurants & Cafés
03
“We cannot have an assistant that invents a fee, a date or a source.”
Answers come only from your own records, link the source, and say ‘I don’t know’ otherwise.
Seen in: Real Estate, Law, Tax & Accounting Firms, SaaS & Startups
04
“Requests arrive as free text and someone retypes them.”
Code handles the fixed rules, AI turns the free text into clean fields, and a person approves before anything is saved.
Seen in: Logistics & Freight, Manufacturing & Export
05
“We were told to do AI, but our data cannot go into a public chatbot.”
First we decide where it is hosted, which AI provider to use, how long data is kept and what is logged. Only then is any data loaded, and AI tools get access that is recorded and checked.
Seen in: Law, Tax & Accounting Firms, SaaS & Startups
What we build
Five systems we put into production, and one plan for deciding which
Customer Assistants
An assistant that answers customers from your own catalogue, policies and order status, and hands over to a person instead of guessing.
Retrieval Over Private Data
Staff get answers from your own contracts, manuals and support tickets, with a link to the source. Each person sees only what they are allowed to.
Agent Pipelines
AI reads what comes in and takes the next steps in your systems, such as turning an emailed order into a draft order. Every step is logged, and a person approves where needed.
MCP Servers
An MCP (Model Context Protocol) server lets AI tools such as Claude work with your app, store or API. They can only take the actions you allow, and every action is recorded.
Process Automation & Integrations
Rule-based connections between systems you already pay for, often with no AI, because that is cheaper to run.
Not sure yet?
AI Consulting & Roadmap
A written plan of which workflows to try first and what each costs to build and run, usable with or without us.
Start With a RoadmapWhat we build with
Anthropic Claude API
OpenAI API
Google Vertex AI
Model Context Protocol
Laravel / PHP 8.4
Laravel Horizon
Redis
MySQL
Meilisearch
Livewire
Next.js / TypeScript
Expo / React Native
How a job starts here
One conversation, then a diagnosis, before anyone quotes
Free technical meeting → System Review, credited to the build → fixed-price Build → monthly Run.
Start: What You Bring Us
In the free technical meeting, bring one real example of the work, where its data lives and who may see it.
Check: The Task Before the AI
The System Review checks the task, the data and how we would know an answer is correct. Often the honest answer is a simple rule or a basic link between two systems, and the report says so.
Build: One Task at a Fixed Price
If the task passes, we choose the right kind of build for who it serves: customers, staff, back-office work or an AI tool. It is built at a fixed price agreed before any code is written.
Finish: What You Keep
It runs in your own code repository and accounts, with its test questions and activity logs.
Once the work is agreed: study the work, not the hype, then one task live first, then measure and make it safe, then run and extend.
How we think
Three rules behind every AI system we ship
01
Advice comes before code
AI in the wrong place adds cost and delay to every request, and a new way to be wrong. Where a simple rule or a good search box works better, we say so.
02
Most of an AI system is not AI
A working AI system is mostly ordinary software: task queues, access rules, checks and logs. We use AI only where something must be read or written, because plain code is cheaper and easier to test.
03
The model is a replaceable part
We build on the Anthropic Claude API, the OpenAI API or Google Vertex AI behind one shared layer. When a provider retires a model, switching is a settings change and a test run.
In production
We run AI products of our own. That is where the advice comes from
FAQ
Questions people ask before AI work
For specific tasks, yes; as a general goal, rarely. It pays where your team spends hours weekly reading, sorting or drafting, not where a rule would do.
It depends on the task. Your first live AI task gets a fixed price before any code is written. We also estimate the AI usage fees with you up front. Current prices are on our pricing page.
Often, yes. For staff writing or research, a team subscription is cheaper. Custom work pays off when AI must use your data, work inside your systems or talk to customers.
It stays in your systems. Only the information a request needs is sent to the AI provider, and we agree on that limit before we build.
