How an AI feature fits together
The model is one part. Most of the work is giving it the right information, limiting what it can do, and checking what it produces.
- You have
Your documents and systems
Policies, records, tickets and the systems your team works in.
- Step 1
Search and tools
Finds the right information and the actions it is allowed to take.
- Step 2
Assistant
Answers or acts, within the permissions you set.
- Step 3
Checks and review
Every action is logged, and unclear cases go to a person.
AI development services
Every AI feature is measured against a set of test questions before launch, runs in your cloud account, and passes anything it is unsure about to a person.
Assistants
- Assistants that actLook things up and act in your systems, with approval for anything that cannot be undone.
- Answers from your documentsAnswers based on your own documents, with a link to the source.
- Voice and chatPhone and chat agents that work from real data and hand over to a person.
- MCP connectionsYour systems made available to AI tools once, with the same access rules everywhere.
Documents and data
- Document readingInvoices, forms and statements read into structured data and checked against the source.
- Review queuesAnything that fails a check goes to a person, with the evidence attached.
- Machine learningForecasting, classification and scoring where a language model is the wrong tool.
- MonitoringQuality, cost and response time tracked after launch.
Tools we work with
We choose the model per feature, based on the task, the cost and your data rules.
Models
- Amazon Bedrock
- Model vendor APIs
Search
- Pinecone
- Weaviate
- FAISS
- pgvector
Machine learning
- scikit-learn
- XGBoost
- PyTorch
- SageMaker
Connections
- MCP
Questions we get on the first call
Have a question that is not here? Send it with the form below and an engineer will reply within one business day.
Which models do you use?
The one that fits the task, the cost and your data rules. In practice that means models on Amazon Bedrock or directly from the model vendors, chosen per feature and easy to swap.
Is our data used to train anyone’s model?
No. We use models under commercial terms that exclude training on your data, keep search indexes in your own account, and can run inside your private network where the provider supports it.
How do you stop the AI from making things up?
Answers are based on your documents with a source for each, anything factual comes from your data rather than the model’s memory, and a test set measures how often an answer is unsupported. Cases that fail go to a person.
What does it cost per request?
We estimate it before building, based on the model, the amount of text and the number of lookups. Spending limits are enforced while it runs, so a bug cannot run up a bill.
Can it run in our own cloud?
The search, the assistant and its tools run in your cloud account. Where the model runs depends on the provider: Bedrock keeps traffic inside AWS, and some open models can be hosted by you.
How long does a first version take?
It depends on the use case and how quickly we get access to your data. Agreeing the scope and building the test set usually take the first two weeks.
Four ways to work with us
A defined project
A defined integration or product, priced and scheduled up front, delivered with the test suite and infrastructure code you keep.
Engineers on your team
Engineers who join your team, your code and your daily meetings for as long as the roadmap needs them.
Ongoing support
We keep the integrations you run working: monitoring, vendor updates, new connections and month-end support.
Taking over existing systems
A system or integration nobody wants to touch: we review it, make it stable, then build on it.
