Kearney · Consulting · Digital transformation
A diagnostic platform for digital transformation, with AI in the loop
A continuous-improvement platform that processes enterprise data to surface pain points and generate improvement opportunities — with a human validating every output before it counts.
- Role
- Sole integration owner and last technical authority
- Period
- 2026 — live
- Stack
- Django · PostgreSQL · React · TypeScript · Docker · Azure · Claude
The problem
A transformation diagnostic is a large pile of unstructured evidence gathered across an organisation. Turning it into findings is slow, inconsistent between consultants, and almost impossible to audit afterwards.
The platform runs the methodology as software and uses language models for the synthesis steps a consultant would otherwise do by hand.
Architecture
Django and PostgreSQL behind a REST API, React and TypeScript on the front, containerised and delivered as self-hostable software, deployed on Azure. Live with the client and in use through the 2026 engagement period.
Self-hostable mattered more than it sounds. Diagnostics contain exactly the material a client will not put in someone else’s cloud, so the deployment model shaped the architecture from the start.
Where the models sit
The models read from the accumulated evidence in the tool rather than from a single prompt, and produce the synthesis layer: the structured read of what the organisation said about itself, and the improvement opportunities that follow from it.
Getting several different kinds of evidence into one coherent input is the hard part. Generation is not.
Human in the loop, by design
The model does not have the context a consultant has — the politics, the thing said off the record, the client’s actual appetite for change. So nothing it produces is actionable until a person accepts it. Promotion from proposed to real is an explicit step, never a default.
That design has a useful property: promotion is a recorded accept/reject judgement on model output, which is an acceptance rate waiting to be measured. The engagements are early enough that there is no number yet. When there is, it will be the most honest quality signal the platform has.
Evaluation
The weakest part, and worth stating plainly. There are no automated evals and no regression tests on model output. What exists is qualitative: consultants report the synthesis is better than what they would have written and far faster.
There is production observability — activity, AI usage and errors are monitored in-app — so failures and cost are visible. Quality is not. Knowing the difference between those two matters: monitoring tells you a system is running, not that its output is any good.
Building it
Three people, of whom two are consultants rather than engineers, and I am the sole integration owner and last technical authority.
The team was never going to be three full-time developers, so the workflow had to absorb that. Everything is built with a specs-first, multi-agent method covering implementation, review, and infrastructure deployment and diagnostics. The outcome I care about more than the speed: non-developer consultants were onboarded into the process and contribute to a real codebase.
What I would do differently
Expect adoption to lag capability. The feature that would prove the AI works is the one users have not reached yet. Shipping capability ahead of workflow readiness has a real cost and I underestimated it.