Four engagements, described honestly
Client names are withheld under our standard confidentiality terms, and we have not attached invented percentages to outcomes we cannot publish. We are happy to talk any of these through in detail on a call.
Demand forecasting for a logistics operator
Replanning ran on a spreadsheet that three people understood and nobody trusted.
Stockouts on the top-moving lines fell materially, and planning moved from three days a month to an afternoon.
What we did
- Baselined the existing forecast error before proposing anything — it turned out to be better than the team assumed, which changed the target
- Built a feature pipeline covering seasonality, promotions, weather and lead-time variability
- Shipped a gradient-boosted model with per-SKU confidence intervals, so planners could see where to trust it
- Deployed with drift monitoring and a monthly retrain, plus a fallback to the old method if inputs go stale
Retrieval over two decades of technical documentation
Engineers were rediscovering answers that already existed somewhere in a 40,000-document archive.
Measured answer accuracy against a held-out question set before rollout, and again monthly afterwards.
What we did
- Ingested and chunked mixed-format archives — PDFs, scans, wiki exports — with OCR where needed
- Built a hybrid retrieval layer combining lexical and semantic search, evaluated against a hand-labelled question set
- Every answer cites the source document and page, so an engineer can verify before acting
- Access control inherited from the existing identity provider rather than reimplemented
Cloud cost engineering for a SaaS platform
Cloud spend was growing faster than revenue and no team could say which feature caused it.
Run-rate fell substantially and, more usefully, stayed down — the guardrails outlast the exercise.
What we did
- Built cost attribution down to service and customer tier, which alone changed three roadmap decisions
- Rightsized compute and storage, and moved batch workloads to interruptible capacity with checkpointing
- Set commitment coverage against a modelled baseline rather than last month’s peak
- Added budget guardrails and anomaly alerts so a regression is caught in days, not at invoice time
Lakehouse and governance for a healthcare analytics firm
Analysts had data but no defensible story about where it came from or who could see it.
The governance model was the deliverable the client valued most — it was what unblocked their enterprise deals.
What we did
- Landing zone and account structure rebuilt with identity, network and guardrails set before workloads moved
- Lakehouse with an open table format, versioned datasets and lineage from source to dashboard
- PII classification, masking policies and retention rules enforced by the platform
- Evidence for audit generated from the platform rather than assembled by hand each quarter
Ask us about any of these
Including the parts that went badly. Book a call and we will talk through the approach, the numbers and what we would do differently.