Services
Every engagement is fixed-scope and fixed-price, with the evaluation criteria written down before any code is committed. If we can't define what "working" means, we don't start.
Service 01
Two to four weeks. For organisations with a budget approved and no agreement on where it should go.
Most AI shortlists are assembled from what is visible rather than what is valuable. We rebuild the list from the process side: where the time actually goes, which of those steps a model can genuinely do, and what it would cost per transaction at your volumes.
You leave with a ranked roadmap, a costed recommendation, and a working spike against the riskiest assumption.
Service 02
Typically eight to fourteen weeks. Full delivery of an AI product against your data and your workflows.
This is the bulk of what we do. Retrieval over internal knowledge, extraction and classification of documents at volume, drafting and review tools, and natural-language interfaces onto systems your team already lives in.
Built as ordinary software: version-controlled, tested, traced, deployed through your existing pipeline. The model is one component in an architecture, not the architecture.
Hybrid search across documents, tickets, wikis and databases, with the chunking, reranking and permission filtering that makes answers accurate and access-safe.
Extraction of structured fields from contracts, invoices, forms and correspondence, with confidence scoring and a review queue for anything below threshold.
Generation grounded in your own precedent and house style, with the editing surface built for the people who will actually use it every day.
Question-answering over operational data, with generated queries validated before execution and results traceable back to source records.
Service 03
Systems that take action rather than produce text — and the guardrails that make that acceptable to your risk function.
Agency is a spectrum, and most vendors sell the far end of it. We start from the least autonomy that solves the problem and add more only where it earns its keep. A state machine that calls a model at three decision points is easier to trust, cheaper to run, and simpler to debug than an open loop.
Where genuine autonomy is warranted, it ships with scoped credentials, spend ceilings, rate limits, complete audit trails, and human approval on anything with a blast radius.
Service 04
For teams who already built something and now need it to be dependable. Usually four to eight weeks.
You have a system in front of users. It mostly works. Nobody can say by how much, quality moves unpredictably when anyone touches the prompt, and the model bill is climbing faster than usage. This is the engagement for that.
Tracing on every call, cost and latency attributed per feature, and failure modes categorised from real traffic rather than assumption.
A graded test set drawn from your own production data, so any future change can be scored before it reaches users. This is the artefact teams most often lack and most need.
Caching, model right-sizing, prompt compression, batching. Reductions are measured against the eval suite so savings never come at silent expense of quality.
CI that runs evals on every pull request, alerting on drift, documented runbooks, and pairing with the engineers who will keep it alive.
Working together
Fixed price against a fixed scope, quoted after a short paid discovery or a free scoping call, depending on how well-defined the problem already is. You will know the total before we begin. If scope changes mid-engagement we re-quote the change rather than quietly absorbing it into a day rate.
Yours. Your cloud accounts, your model provider accounts, your repositories. We do not host your system on a platform of ours, and there is no runtime you have to keep licensing. If you end the relationship, everything keeps running exactly as it did.
Whichever is right for the task, and we keep the swap cheap. In practice that means frontier models where genuine reasoning is required, smaller and cheaper models for classification and routing, and an abstraction thin enough that changing provider is a configuration decision rather than a rewrite.
Usually. We routinely work inside VPCs, with zero-retention model endpoints, regional processing requirements and existing permission models. Bring your security team into the scoping call — it is considerably cheaper to design around their requirements than to retrofit them at the end.
Your team owns it. Handover includes documentation, architecture walkthroughs and pairing sessions, so ownership is real rather than nominal. Retainers for ongoing development and support are available if you want them, but nothing about the build is designed to make you need one.
Regularly. If the success criteria can't be measured, if the data needed doesn't exist yet, or if a deterministic solution would be cheaper and more reliable, we will say so — usually in the first conversation, before you have spent anything.
Next step
Describe the process you want to change and we'll tell you whether AI is the right instrument, roughly what it would cost, and what we would build first.