Cases, credentials and artefacts an auditor would recognise.
Three anonymised production cases closest in kind to the work we propose. Credentials that get vetted before the work starts. And samples of the artefacts we produce during delivery — not marketing collateral, the real thing.
Credentials
Three cases, in production
Anonymised. Each in the same schema — context, what we built, what it measurably changed, what stayed with the client. Named contacts on request.
Context
Authorised EU financial services firm, DORA Article 28 in scope, AI in the operational chain.
What we built
AI control layer over the production system: per-request audit trail, guardrails, exit-readiness pack.
Measured result
Passed audit review with zero compliance issues. Eight weeks to production.
What the client kept
The source code, the accuracy tests, the alert catalogue and the DORA-aligned artefacts — in the client's repository, ready for their next audit.
Context
Several operational systems each held part of the picture in whatever shape each system used. Products the client wanted to launch were not feasible while the data stayed fragmented.
What we did
Pulled the same entity out of every system, agreed one shared description the whole organisation could work from, and kept the source systems in place.
Measured result
Products the client subsequently launched — and that had not been possible while the data stayed fragmented — went live on that platform.
What the client kept
The shared description, the pipelines that keep it accurate, and the rules that resolve disagreements — running under the client's operations team.
Context
Data pipelines feeding machine-learning models at volumes where nobody can spot-check the output. The failures that hurt are the ones that look plausible and are wrong.
What we did
Built the layer that watches pipelines and models for exactly that failure mode — the one where the dashboard is green and the answer is wrong.
Measured result
Model behaviour became traceable to the exact data behind it. When a number moved, the team knew whether the model changed, the data changed, or something else.
What the client kept
The measurement discipline. Silent failure stopped being invisible and started being a category the team owned.
Context
Built from scratch, ordinary in kind, demanding in scale and setting — volumes at which an inefficient transformation is unaffordable rather than merely slow.
What we built
Pipelines processing terabytes into the warehouse. Every design decision documented, reviewed and defended before it was built.
Measured result
Deliverable at professional-services quality bars, with an audit trail on every architectural choice.
What stayed
The warehouse, the pipelines and the review record — the standard the client kept using.
On request: a divisional reporting platform at PepsiCo, and the architectural rebuild of a production LLM platform where the technical team was led into a target architecture that cut deployment time and maintenance cost.
Sample artefacts
The artefacts we produce during delivery, in the shape they are actually delivered in. Anonymised. Fragments only — enough for your architects to know what to expect.
Architecture annex page — with [assumption] markers
One page from an architecture annex, showing how we mark assumptions and name the section that closes each one. The gap we are not dressing up.
Evaluation harness report — with reliability diagram
Per-field accuracy, calibration reporting, cost-per-request-type reported alongside accuracy. What the harness looks like at the pilot gate.
Compliance pack — table of contents
The regulator-, auditor- and works-council-ready artefacts we produce during delivery. GoBD Verfahrensdokumentation section, AI Act classification, DORA ICT-register entries.
Book an architecture review.
45 minutes. One process, one obligation, one system-of-record boundary — mapped honestly against your evidence, not against a demo.
Book a review