Services
AI agents, AI products, retrieval systems, and the software around them.
We design, build and support software with intelligence built in, to the same engineering standard as everything else you run. Each engagement starts with a two-week pilot, a fixed price and a working demo.
01 AI agents
Agents with a defined job and a clear log
An agent is a model that can call tools: search, query a database, call an API, draft a document. A multi-agent system splits a larger job between specialists. We keep the job narrow and the behaviour inspectable.
What we build
- Single agents with tools for your APIs, databases and documents
- Planner, worker and reviewer patterns where one agent is not enough
- Human approval steps before anything that changes data or sends a message
- A trace of every step, so a run can be replayed
How we know it works
- A task suite that scores end-to-end success, not single answers
- Cost and latency per run, tracked over time
- Regression tests that run when a prompt or model changes
First step. One workflow, one agent, one measurable outcome, built and evaluated in a two-week pilot.
02 AI products
AI products, from first prototype to paying users
Some companies want AI inside the software they already run. Others want a new product built around it. We take a product from a tested idea to a launched release, then keep improving it with real usage data.
What we build
- Discovery: the problem, the users, and what a first version has to prove
- A prototype in front of real users within weeks
- A production MVP with accounts, billing, analytics and admin tools
- AI features designed for cost, latency and failure from the start
How we know it works
- Success criteria agreed before we build, and tracked after launch
- An evaluation set for every AI feature, run on each release
- Usage, cost and quality dashboards you can read without us
First step. A two-week discovery and prototype, ending in a demo and a fixed-price plan for the MVP.
03 Retrieval systems (RAG)
Retrieval systems that cite their sources
Retrieval-augmented generation (RAG) lets a model answer from your own documents instead of from memory. We build it so every answer points to the passage it came from.
What we build
- Ingestion for PDFs, wikis, tickets and shared drives, with chunking tuned to your documents
- Hybrid search (keyword and vector) with reranking
- Answers with citations to the exact passage
- Access control that follows your existing permissions
How we know it works
- An evaluation set built from your team's real questions
- Retrieval hit rate and answer faithfulness, measured before and after each change
- A list of questions the system declines to answer, and why
First step. A two-week pilot on a slice of your documents, with a working demo you can test.
04 Software engineering
Software engineering, to production standard
Most of a working AI system is ordinary software: interfaces, APIs, data pipelines, permissions and infrastructure. We build all of it, and we build software with no AI in it to the same standard.
What we build
- Web applications, dashboards and internal tools
- APIs, back ends and integrations with the systems you already run
- Data pipelines, cloud infrastructure and CI/CD
- LLM integration: a model gateway, structured outputs, redaction, spend caps and fallbacks
How we know it works
- Code review on every change, against written engineering guidelines
- Automated tests in CI, with quality gates that block a release
- Security checks, dependency updates and a runbook for whoever runs it next
First step. A short technical review of your codebase or plan, then a fixed-scope first milestone.
Tell us what you want your software to do next.
A 30-minute call is enough to say whether we can help, what it would take, and what a two-week pilot would look like.
Book a call