No numbers on this page, on purpose.
We would rather show you the shape of the work than a percentage you cannot check.
The results a consultancy publishes are usually the ones that survived selection — the engagement that went well, the metric that moved, the baseline chosen after the fact. You have read enough of them to discount them automatically, and you are right to.
Our largest engagement is under an NDA that covers its figures, and the numbers we could quote from it have not been independently verified. So we do not quote them. Publishing a number we cannot evidence would cost us more than the blank space does, on a site whose whole argument is that we will tell you what does not work.
What we can describe is the shape: what the problem was, what turned out to be hard, and what we would do differently. That is the part you can reason about anyway.
If you want references, ask during a conversation. We will arrange them where the client has agreed to it, and say so plainly where they have not.
Enterprise healthcare — clinical operations
A large healthcare organisation with a document-heavy operational process, a regulated data boundary, and an engineering team that had not previously shipped anything built on language models.
What made it hard
The data could not leave a defined boundary, which removed most of the obvious architecture before the project started. The review process that made the output trustworthy was also the thing that made it slow, so removing it was not on the table — the work had to make review faster rather than optional.
And the failure mode that mattered was not the system being wrong. It was the system being confidently wrong in a way a busy reviewer would pass, which is a design problem long before it is a model problem.
What we would do differently
Spend longer on discovery. We took the documented process as the specification for longer than we should have, and the gap between it and the real one surfaced during the build rather than before it. That is the reason discovery is now written into how we scope rather than treated as preamble.
On the numbers
Covered by the agreement, and unverified on our side. We are not going to publish them here. In a conversation we can describe what improved and how it was measured, within what the NDA allows.
The more useful test.
Describe a problem you are weighing and see whether the answer you get is specific, and whether it includes the parts that would not work. That tells you more about a firm than any case study does — including this one.