AI is not a product we sell you.
It is a layer we wire in.
Your ERP, your document store, your ticketing system, your data warehouse — they already hold the answers. We build the retrieval, orchestration and evaluation layer that lets a model use them safely, and we ship it into the product your customers already use.
Left: what you already own. Right: where it has to show up.
The interesting engineering is in the middle column. Anyone can call a model API — the work is making it answer from your data, under your permissions, with a way to prove it was right.
Runs in your cloud tenancy
Where each module actually stands
Ordered by maturity, not by how impressive it sounds. “Live” means it is running against real data for a paying client. Every row also states what it is still bad at — because that second column is how you judge the first.
Grounded document assistant
Answers questions over your contracts, policies and runbooks, citing the paragraph it used. Refuses when retrieval comes back empty.
Strong on finding the deciding clause. Still weak on documents that contradict each other — it flags the conflict rather than resolving it.
Service desk triage
Reads an inbound ticket, classifies it, attaches the customer's entitlement and history, and drafts the first reply for an agent to approve.
Reliable on the top twenty intents. We deliberately route the long tail to a human instead of guessing.
Structured extraction
Turns invoices, purchase orders and scanned forms into validated records, with confidence scores and a review queue for anything below threshold.
Accuracy is good on clean scans and honest about the rest — low-confidence items go to a person, not into your ledger.
Product imagery generation
Generates on-brand product and lifestyle imagery from a controlled style reference, with a human approval step before anything publishes.
Not usable yet for products with fine text or precise logos on the item itself.
Agentic back-office workflows
Multi-step internal processes where an agent takes permissioned actions across systems, with every step logged and reversible.
The interesting part is not the agent, it is the authorisation model. We scope each agent to a short, explicit list of write actions.
Voice intake
Handles routine inbound calls — order status, appointment changes — and hands off with the transcript attached.
Still in design. We are not shipping a voice agent until barge-in and interruption handling feel genuinely natural.
Five rules that do not bend
These apply to every module on the board, including the ones still in design. They are the reason we can put an AI system in front of an auditor.
Retrieval before generation
If a claim can be looked up, it gets looked up. The model composes the sentence; it does not invent the fact.
Permissions are the user's, not ours
Retrieval runs as the signed-in user. There is no service account that can see everything.
A short list of write actions
Anything that changes your data is an explicit, authorised, logged action — never an inference.
Evaluation is a build artefact
A graded test set ships with the system and runs on every change. Regressions block the release.
Model-portable by default
Providers are configuration. Swapping a model is a deploy, not a rewrite — so you are never hostage to one vendor's pricing.
Start small enough that being wrong is cheap
AI readiness review
Two weeks. We inventory your data, permissions and processes, then tell you which three use-cases are worth funding — and which ones are not.
Six-week pilot
One intent, your data, behind a feature flag, with an evaluation set you keep. Cancellable at the halfway point.
Embedded AI engineers
GenAI and ML engineers on our payroll, inside your sprints, reporting to your lead. Scale up or down monthly.
Bring us the process, not the prompt
Tell us which workflow is slow, expensive or error-prone. We will tell you whether AI is the right tool for it — including when the answer is no.