Selected work · 2021 — 2025

0 engagements shipped

Systems in production,
with the numbers to prove it.

Every case below is a real deployment running real traffic. We publish the constraint we were handed, the system we built and what changed once it went live. Names of clients under NDA are described by sector and scale instead.

02 Freight & supply chain

Kestrel Logistics

An agentic dispatch layer that handles exceptions, not just orders.

Kestrel's dispatchers were re-planning by hand every time a truck was delayed. We built a bounded agent that watches telemetry, proposes reroutes against cost-and-service constraints, and files the decision — with a full audit trail — for a human to approve or override. It never books a carrier without a signature.

  • 0%of exception reroutes proposed automatically
  • 0 minmedian dispatch decision time, down from 47
  • 0autonomous bookings without human approval
Agentic workflow Human-in-the-loop 6 weeks
03 Clinical operations

Verity Health

A scheduling model that respects nurse ratios, not just calendar gaps.

Verity's rostering tool optimised for empty slots and quietly broke staffing ratios in the process. We replaced the objective from the ground up: a constraint-solver wrapped in a forecast of expected admissions, exposed through the tool their coordinators already used. No new interface, no retraining week.

  • 0%reduction in ratio breaches per month
  • 0%fewer agency shifts booked
  • 0 siteslive, with two more in rollout
Decision model Constraint solving 10 weeks
04 Arcstone Manufacturing · Industrial

Four models, three plants, one on-call rota — and nobody left holding the pager alone.

Arcstone had accumulated four separate AI pilots across two years: a defect classifier, a demand forecaster, a maintenance predictor and an invoice extractor. Three of them worked. All four had been abandoned by whoever built them, and by 2024 nobody on staff could safely change any of them.

We took over the whole estate. That meant consolidating the deployment pipeline, writing the runbooks nobody had written, adding drift and data-quality monitors, and putting every model behind the same evaluation harness so a change to one couldn't silently break another. The engagement is ongoing — we hold the pager one week in four.

0 models consolidated onto one pipeline
0p99 latency held within budget across all endpoints
0 unmonitored models remaining in production
0 runbooks written and rehearsed with plant staff

Managed AI operations is the work almost nobody wants to sell you, because it never looks like a launch. It is, in our experience, the difference between a model that is still running in year three and one that got quietly switched off.

Read about managed AI operations

What these engagements have in common

None of the projects above started with a model choice. They started with a concrete operational bottleneck and a number we agreed to move. The technology followed from that.

  • We wrote the failure path first

    Refusal, escalation and fallback behaviour was specified and tested before the happy path existed.

  • Every claim traced to an evaluation

    The metrics on this page come from live monitoring, not a run of a hundred hand-picked samples.

  • Nothing shipped without an owner

    We hand over with a named internal owner, a runbook and a rehearsal, or we stay on to operate it.

  • Scope shrank, not stretched

    Each project began smaller than the brief, proved the case, and then earned its expansion.

Tell us the bottleneck.
We'll tell you if a model fixes it.

A 30-minute call, no deck. Bring the process that is costing you the most time and we will say plainly whether AI is the right lever — and if it isn't, we will say that too.

See how we work