AI Development
AI development services.
We add AI to the product you have, or build a product with AI at the core. That work spans assistants and other LLM features, answers grounded in your own data, and agents that carry out work inside your product. Every feature is measured against real examples before it ships.
01 · Two ways in
Add AI to your product, or build one around it.
Most AI work begins inside a product that already exists. Some products are only possible because of what AI can do. We do both.
Add AI to an existing product
You have a product, users, and a list of things AI could do for them. It could answer questions from your own documentation, summarize, draft, classify, or carry a multi-step task through from start to finish. We identify the feature with the clearest value, build it into your product, and measure it before it ships.
Talk about your productBuild an AI-native product
The idea only works because of what AI can do. We shape the product around that from the start, deciding which model does what, where the data comes from, what happens when the model is wrong, and how the product stays useful as models change.
Talk about your idea
02 · Our AI development services
We offer five kinds of AI work, each built into web and mobile products and measured before it ships.
AI assistants and chatbots
We build customer-facing chatbots and in-product assistants that answer questions, guide users and hand over to a person when they should, built with Claude, GPT, Gemini or open models, chosen per use case.
Generative AI features
Summarization, drafting, classification and extraction are built into the flows your users already have, from the first draft of a reply to the tags on a ticket to the five fields pulled out of an uploaded document.
Answers from your own data
We build retrieval over your documents, tickets, records or knowledge base, so the assistant answers from what your company actually knows and shows where each answer came from.
AI agents and workflow automation
We build agents that carry out real work inside your product or your operations. They call your tools and APIs, run multi-step workflows, and pause for a person’s approval at the steps you decide should not run alone. We build them with tool calling and MCP where it fits.
AI integration into existing products
We add the right AI to the product you already run, guided by your needs, your data and your budget, and build it in a way your existing team can maintain.
03 · How we build AI features
Four steps, with a measured result at the end of each.
Frame
First
The use case with the clearest value, and what a good answer looks like.
A short working session identifies the use case with the clearest value, defines what a good answer looks like, and confirms what data is available to work with. If AI is not the right instrument for the problem in front of you, we advise you accordingly.
Prototype with real data
Before the build
A prototype answering from your real data, and the evaluation set to judge it by.
We build a working prototype against your real data, so that you and your users judge actual answers rather than a demonstration. An evaluation set built from real examples takes shape at the same time, so that quality is measured rather than estimated.
Build and integrate
Through the build
The feature in your product, measured on every change.
The feature enters your product with the same discipline as any other. Automated tests and the evaluation set run on every change, a second engineer reviews the changes that carry risk, and monitoring is in place before launch. Your data remains in your own accounts and is not used to train anyone’s models.
Most often
- Claude
- GPT
- Gemini
- Open models
Launch and monitor
After launch
Answers that stay measured as the models and your data change.
Models change, data changes, and answers drift over time. Monitoring remains active after launch and the evaluation set runs on every change, so a shift in a model or in your data appears as a number before it reaches your users.
04 · Three rules on every AI project
Whatever the model and whatever the product, these three principles hold.
Model-agnostic
We work with Claude, GPT, Gemini and open models, and we select per use case on quality, cost and where your data is permitted to go. When your data cannot leave your environment, we run open models inside it.
Your data stays yours
Model providers, vector stores and cloud infrastructure run under your own accounts rather than ours, and we apply the settings that keep your data out of anyone’s training.
Measured before it ships
Every AI feature receives an evaluation set built from real examples, and the work is not complete until the numbers confirm it.
05 · What we build with
The stack is chosen for each product.
These are the technologies we work with most.
Models
Claude, GPT, Gemini or open models, chosen per use case.
Retrieval
Vector search in PostgreSQL or a dedicated vector store, whichever the product calls for.
Application
TypeScript and Node.js services inside your existing product, on the web or in the app.
Cloud
AWS first or Cloudflare where it fits, under your own accounts.
06 · Related work
A model inside a product.
Moon Quest forecasts how the days ahead will feel with a prediction model built into the app. It learns from each user as they check in, runs entirely on the phone with no request to a server, and shows the confidence it actually uses. It was the most heavily tested part of the app.
07 · Frequently asked questions
What teams ask about AI.
What AI development services do you offer?
Our services cover AI assistants and chatbots, generative AI features such as summarization, drafting, classification and extraction, answers grounded in your own data through retrieval, AI agents and workflow automation, and the integration of all of these into products that already exist. We build them into web and mobile products, either extending a product you already operate or building one around AI from the start.
Do you integrate existing models or train your own?
We build with existing models. Claude, GPT, Gemini and open models are chosen per use case. Most products progress further and faster with the right model and the right data around it than with a custom-trained one, and that is where we put our effort. We do not train custom models.
Which models do you use?
We use Claude, GPT or Gemini for most features, and an open model when your data cannot leave your environment or when cost matters more than the last point of quality. We select the model per use case and structures the product so that switching models later is a small change rather than a rewrite.
How do you keep our data private?
Everything runs under your own accounts, including the model provider, the vector store and the cloud. We apply the provider settings that keep your data out of model training, and when data is not permitted to leave your environment at all, we run open models inside it.
How do you know an AI feature is good enough to ship?
We measure it. Before anything ships, an evaluation set built from real examples establishes how often the feature answers correctly, and it runs again on every change afterwards. A person reviews the failures and decides what good looks like. We do not ship an AI feature on the strength of a demonstration alone.
Can you add AI to a product you did not build?
Yes. Most AI work begins inside an existing product. We start with a focused review of your codebase and your data, identifies the feature with the clearest value, and builds it in a way your existing team can maintain.
How long does an AI integration take?
The timeline depends on the feature, the state of your data and how much of the product it touches, so we give one after the framing session rather than before it. Adding a feature to an existing product with existing models is the fastest kind of AI work. Building a product around AI from the start takes longer.
Blog
From the blog
How to decide whether to build or buy AI for your product
An AI capability usually takes one of three shapes. It can be an off-the-shelf tool, a feature built on a foundation model, or a custom model. This guide covers how to choose between them on value, data, cost over five years and the risk of being wrong.
RAG explained for founders, making an AI assistant answer from your own data
Retrieval-augmented generation is how an AI assistant answers from your documents and records instead of guessing. What it is, when you need it, where it fails in production, and what it takes to build one that holds up.
08 · Services
Everything else we build.
AI features rarely stand alone. We design and build the product around them.
Build AI into your product.
Tell us what your users need and what data you have available, and we will tell you whether AI helps here and what it would take.