04 · AI engineering
AI that does one job well, instead of five badly
Most AI projects fail for a dull reason: they start with the technology instead of a task. We start with a workflow that costs you hours — reading documents, answering the same question, extracting the same fields — and build only what that needs.
Who this is for
- Businesses reading and re-typing information from PDFs, invoices or forms
- Teams answering the same twenty questions from customers or staff every week
- Owners who have tried a chatbot, watched it make something up, and want it done properly
What is included
- Document extraction and classification — invoices, contracts, forms, reports
- Retrieval systems over your own documents, so answers cite the source rather than invent it
- Assistants scoped to a defined task, with limits on what they are allowed to say
- Evaluation before launch: a test set and a measured accuracy number, not a demo
- Cost controls and logging, so spend and behaviour are both visible
How it works
- 01
Pick the task
One workflow, measured in hours per week. If it cannot be measured, it is not the right first project.
- 02
Prototype
A working version on your real data within two weeks, not a slide deck.
- 03
Evaluate
Accuracy tested against cases you supply, with the failure modes named out loud.
- 04
Ship with limits
Guardrails, human review where it matters, and monitoring on cost and output.
Questions
Prototype-first engagements, scoped after a free discovery call.
Is our data safe?
It depends on choices we make deliberately: which provider, which region, what is retained, and whether anything is used for training. Those choices get written down before the build, and sensitive workloads can run where the data does not leave your tenancy.
How accurate is it, honestly?
For structured extraction from consistent documents, high — and measurably so. For open-ended answering, good enough to be useful and never good enough to leave unsupervised on decisions that matter. Anyone quoting a single accuracy figure without seeing your data is guessing.
What does it cost to run?
Per-use API costs for most workloads, typically tens of dollars a month at small-business volume rather than hundreds. We put a ceiling and monitoring on it before it goes live.
Could we just use ChatGPT?
Often, yes — and we will say so. Paying for a custom build to do what a $30 subscription already does is a bad trade. The work is worth it when it has to run unattended, touch your own data, or connect to your systems.