Selected work
Case notes, not brochures.
Client names are withheld where contracts require it, but the numbers are the ones we reported at handover. Ask us on a call and we will introduce you to the teams who lived through these.
Each note follows the same structure — the problem as the client described it, the approach we argued for, and the outcome we could measure. Where an approach failed first, we say so: the second attempt is usually the interesting part.
01 / Payments
A settlement engine replaced without an outage
Problem
A long-lived settlement monolith processed billions in volume each year and had defeated two replacement attempts. Every release required a weekend window and four engineers on standby.
Approach
We put a strangler facade in front of the monolith, moved one settlement type at a time behind dual writes, and built a reconciliation harness that compared old and new outputs on live traffic for six weeks before any cutover.
11 months
to full cutover
0
unplanned outages
-63%
settlement latency
02 / Logistics
Fleet telemetry that fits in a phone signal
Problem
Nine thousand vehicles reported position and sensor data over patchy mobile networks. The existing pipeline dropped events silently and the operations team had stopped trusting the map.
Approach
An edge buffer with delta encoding, an idempotent ingestion API, and a ClickHouse store with a materialised view per operational question. Dispatchers got a redesigned board built with them over four on-site weeks.
99.98%
event delivery
8x
cheaper storage
4 wks
to dispatcher rollout
03 / Climate
Grid forecasting with an evaluation harness
Problem
A renewables operator wanted machine-learned day-ahead forecasts but had no way to tell whether a new model was actually better than the spreadsheet it replaced.
Approach
We built the evaluation layer first — backtests, error decomposition by weather regime, and a cost function denominated in imbalance penalties — then iterated on models against it in public view.
-19%
imbalance cost
Daily
automated backtests
2
in-house owners trained
04 / Health
Clinical tooling that clinicians defend
Problem
A triage tool designed without clinical input was being bypassed with paper forms in three of five departments.
Approach
Shadow shifts before a line of code, a design system built for gloved hands and bad lighting, and weekly releases reviewed by two practising nurses. Accessibility budgets were treated as failing tests.
94%
adoption in 6 months
-41%
triage time
WCAG 2.2 AA
audited