selected work
Case studies
Three engagements that show how we work: scope in writing, measure before promising, and hand over something the client can run without us. Engagements are anonymised at our clients' request.
Machine learning that has survived five peak seasons
Demand forecasting and depot-level planning support, shipped to production in 2021 and still running today — through every Christmas peak since.
the problem
Planning across depots ran on spreadsheets and experience. It worked until volume spiked, and then it didn't: seasonal peaks meant firefighting, overtime, and missed windows, with no way to test a better plan before committing to it.
what we did
- Started with evaluation, not models: weeks of shadow-mode forecasts scored against what planners actually did before anything influenced a decision.
- Shipped demand forecasting first, narrow and measurable, then extended into depot planning suggestions once the forecasts had earned trust.
- Kept planners in charge — the system proposes, a human disposes. Adoption came from usefulness, not mandate.
- Handed over retraining pipelines and runbooks so the client's own team operates the system without us.
In production since 2021, through five peak seasons
Operated and retrained by the client's own team
Extended twice without a rewrite
A risk engine rebuilt for two orders of magnitude
A recompute path that took 340 milliseconds at p99 was holding real-time pricing hostage. Profiling, a hot-path rewrite, and incremental recomputation brought it to 9 milliseconds.
the problem
Every position change triggered a full risk recompute. At 340ms p99 that meant pricing waited on risk, batch windows crept into the trading day, and adding products made everything quadratically worse.
what we did
- Profiled production traffic before touching code — the slow path was not where anyone assumed.
- Rewrote the hot path against a benchmark suite that gated every change slice in CI.
- Moved from full recomputes to event-driven incremental recomputation, so a position change only recomputed what it touched.
- Delivered before/after performance reports with the numbers, not adjectives.
p99 recompute latency: 340ms before, 9ms after, on production traffic
Risk no longer the ceiling on pricing latency
Benchmark suite left gating the hot path in CI
Figures as published on our services page — representative of the engagement.
From ageing servers to a defensible estate
One of our earliest engagements: a security review that became a staged cloud migration, done around the working week for a firm that could not afford a bad one.
the problem
End-of-life servers under a stairwell, shared logins, backups that had never been restored, and client-confidential work riding on all of it. The firm knew the risk; what it lacked was a path that didn't interrupt fee-earning work.
what we did
- Security review first: an honest written picture of the estate, ranked by what an attacker would actually do.
- Migrated in stages scheduled around the working week, oldest and riskiest systems first.
- Rolled out MFA everywhere, hardware keys for the partners, and retired every shared login.
- Backups rebuilt around tested restores — proven by restoring, not by dashboards.
Ageing on-premises estate retired in stages, no unplanned interruptions to fee-earning work
MFA on every account; hardware keys at partner level
Restores tested and documented, then handed over
The disciplines behind these engagements are on the services page.