B16In production
Packaging concepts, with pixel proof of what changed
Spice brand
Packaging concepts with proof that nothing changed outside the label.
3rendered pages verified
AI guardrails and governance that make a system safe to put in front of customers
Every system we ship runs on four rules: answers come from your approved sources or the system says it has no answer, citations come from the source itself, the working is shown, and a named person approves anything client facing. A manufacturer's system with more than fifteen AI helpers logs every AI action and has a person confirm each one.
Proved in 9 shipped systems
What you get
Answers are grounded in your approved sources, and when nothing matches the system says so, so customers never receive an invented answer.
Citations are taken from the source passage itself, so every reference can be checked.
The working is visible: the reporting chat shows its query and the document reader shows what it read before you confirm.
A named person approves anything that leaves the building, so AI drafts and your team decides.
Every AI action is written to a log that cannot be edited, and agents are built without permissions they do not need, so your audit trail is complete.
Systems we have shipped for this
E02In production
A library that answers questions and always shows its source
Specialist education institute
Answers from approved documents only, each with the page it came from.
65 out of 100minimum match before answering
Live in two weeks
E04In production
A card statement sorter for the accounts team
Company finance
Card statements sorted into company and personal lines, AI only as the last step.
3, cheapest firstsorting steps
First version in two days
E05In production
A WhatsApp admissions helper for parents, built on an audited brief
Private pre-school
Every parent question answered from confirmed facts, or handed to named staff.
16pages reviewed
E10In production
Brand rules written so AI tools follow them
Used across our client projects
One rules file per brand that designers and AI tools both follow.
6 areasrules per brand
E11In production
A factory's complete business system, with more than fifteen AI helpers built in
Transformer and brass parts manufacturer
One system runs the factory every day, with more than fifteen AI helpers.
177+database updates shipped
Shipping weekly releases
E17Internal system
Which video model should you trust? A vendor-neutral benchmark
Internal research
Six video models, one identical input, scored on what breaks in production.
6provider and model combinations
Run and reported in a single day
E19Internal system
An agent that drives a paid tool in the browser, with thirteen safety rules
Internal tooling
An agent that operates a no-API paid tool, with credits protected by design.
13non-negotiable rules
E20Internal system
Costing an agent system before building it
Internal research
Know what an agent system will cost to run before building it.
20 across 8 vendorsmodels priced
What businesses come to us for
Real requests from 2026, with the names taken out.
A promotions agency wanted an AI audit and roadmap, plus policies covering client confidentiality, intellectual property, consumer data, quality control and human oversight.
A consultancy wanted a check layer proving AI generated reports matched the raw source exactly.
A platform team wanted evaluation frameworks confirming AI output was accurate and testable before it reached production.
Guardrails and governance: questions we get asked
- How do you stop AI making things up in a business system?
- We make "no answer" the default. The system searches your sources first, and if nothing matches well enough it replies that it has no answer without calling the AI. Invented answers come from asking a model with nothing to support it, and our design never does that.
- Where should a human sit in an AI workflow?
- On the last step before anything leaves the building or becomes a record: confirming what was read before it saves, approving a script before it is voiced, confirming a payment reminder before it sends. Early enough to be quick, late enough to catch what matters.
- What about prompt injection through documents?
- All retrieved content is treated as information and never as instructions, with a clear boundary in every prompt, so a document cannot redirect the system.
- How do you handle consumer data in AI tools?
- Personal data stays out of general purpose AI tools. Access is enforced per user at the data layer, phone numbers are hidden by default, and every AI action is written to a log that cannot be edited.
- Can you audit AI use across a business?
- Yes, and it is a common starting point. We review your workflows, find the repetitive and error prone ones, rank them by impact against effort and risk, and recommend two or three pilots, delivered as a roadmap you own.
Have a process that should run itself? Let us build it.
Tell us the problem. We will show you what we would build, how fast we can ship it, and what it will deliver, from a first pilot to a full platform.