Services · project or retained
AI built for production, not for the demonstration.
Grounded agents, retrieval, structured extraction and the evaluation harnesses that keep them honest — shipped into the workflow, instrumented for cost, and measured against outcomes that existed before the project did.
The demo is the easy ninety per cent.
A convincing prototype takes a competent engineer a week. The remaining work — grounding the output in data that changes, handling the responses that do not parse, catching the moment the model invents an identifier that does not exist, keeping cost predictable at volume, and proving a prompt change did not break something — is where the actual engineering lives.
Skipping it produces systems that work impressively in review and erode quietly in production. Users encounter a confident, wrong answer, tell three colleagues, and adoption never recovers. The failure is rarely dramatic enough to trigger a fix; it just becomes a tool nobody opens.
We build for the ninety per cent. That means defensive parsing, validation against your own data, deterministic arithmetic, metered spend and a regression suite — from the first version, not added after the first incident.
What we build
What we build
Grounded analysis pipelines
Sources normalised into labelled, provenance-tagged, budget-capped evidence; a strict output contract; and tolerant parsing that salvages a truncated response rather than discarding the work that produced it.
Agents with constrained tools
Tool-calling loops where the model composes safe, structured queries against curated datasets. It never writes executable code and never reaches a data surface it was not given. Flexibility lives in the data layer, not in what the model is permitted to run.
Retrieval, written directly
Embedding, chunking and ranking tuned to your corpus, without a heavyweight framework in the middle. Roughly two hundred lines you can read, debug and adjust — not an abstraction stack you have to reverse-engineer.
Structured extraction
Documents to typed records, with every extracted value locatable in the source page it came from. Anything that does not resolve against your data is dropped rather than passed downstream to be discovered later.
Evaluation harnesses
A regression suite over real examples with a graded rubric, so changing a model or a prompt produces a number. This is the single highest-value artefact of most engagements and the one most often skipped.
Cost and provider control
One choke point for every model call, tagged by feature, with real token cost recorded per request. Cheap triage routing ahead of expensive analysis, and a provider interface that makes moving vendors or moving into your own tenancy a configuration change.
What changes
What this changes
Output people are willing to act on
Citations, an explicit basis, and an honest "insufficient evidence" where that is the truth. Users trust a system that admits uncertainty considerably more than one that never does.
Model upgrades stop being frightening
With an evaluation harness, moving to a newer model is a measured decision made in an afternoon rather than a change nobody wants to authorise.
Spend you can explain
Per-feature cost, visible in the product. When someone asks what the AI is costing, the answer is a breakdown rather than a monthly total and a shrug.
Maintainable by your team
Boring, well-documented technology chosen deliberately. Your engineers should be able to read the code and change it without having been present when it was written.
Bring the workflow, not the model preference.
The right technology choice falls out of the constraints — data residency, latency, cost per decision, who has to defend the output.