Case studies
What the work
actually looks like.
Where clients have published figures, we cite them. Where they have not, we describe the shape of the work rather than invent numbers.
Kron Health: a digital clinic platform
The problem
Delivering a clinic platform where AI-assisted workflows had to fit clinical reality: careful handling of patient information, review by a clinician at the right point, and behaviour that is auditable rather than opaque.
What we did
- Product & platform engineering across clinical and patient-facing workflows
- AI-assisted documentation and intake with clinician review in the loop
- Evaluation and monitoring built in from the start, not retrofitted
Outcome
A shipped, working product, and the clearest illustration of the depth of build work our team takes on when a domain demands care rather than speed.
Cutting manual review on document-heavy onboarding
The problem
An operations team spent most of its day extracting the same fields from inconsistent documents, then re-checking them, with a growing backlog and no way to see where errors originated.
What we did
- Mapped the review workflow and built an evaluation set with the reviewers themselves
- Extraction with confidence thresholds that route uncertain cases to a person
- Integrated into the existing case system so reviewers never left their workflow
Outcome
Reviewers moved from transcription to judgement on exceptions, with an audit trail on every automated decision.
Scaling catalogue operations without scaling the team
The problem
Content operations could not keep pace with new supplier ranges. Product data quality varied by supplier, and inconsistent listings were suppressing conversion.
What we did
- Normalisation and enrichment pipeline over supplier feeds
- Generated copy and attributes held to a house style guide, with merchandiser approval before publish
- Quality measures reported per supplier so the worst inputs got fixed at source
Outcome
A catalogue pipeline the merchandising team runs itself, with human approval kept where brand risk actually sits.
Restarting a stalled quality-inspection pilot
The problem
A promising inspection model had sat unused for two quarters. Nobody on the line could act on its output, and it had never passed internal security review.
What we did
- Diagnosed the stall as integration and ownership rather than model quality
- Deployed inside the plant's own environment with monitoring and continuous evaluation
- Surfaced results in the tool operators already used, and named an owner for the decision
Outcome
The capability moved from a parked notebook to a routine part of the shift, with the plant team owning it after handover.
Start here
Tell us the problem.
If your work involves repetitive judgement over messy data inside systems people already use, there is probably a version of these here for you.