Applied AI
Built for the work, not the demo.
Dependable AI systems running inside real operations, measured on real outcomes.
4,120
Workflows in production
38
Manual handling removed
99.2
Completed without human touch
11
Map to measured production
Practice
Four layers. One operational system.
Applied AI
Models selected, evaluated and constrained for one job inside your operation.
We choose the smallest model that clears the accuracy bar, wrap it in retrieval and guardrails, and keep a human review point on anything customer-facing. A model is a component, not the product.
Low confidence returns to the reviewer with the source attached.
- Evaluation harness
- Retrieval layer
- Review loop
Agentic Systems
Agents with defined scope, tool access and a stop condition you can inspect.
Every agent gets a bounded task, an auditable tool surface and a written escalation path. No agent runs without a cost ceiling, a step ceiling and a way to hand back to a person.
A step or cost ceiling stops the run and pages the owner.
- Task scope
- Tool surface
- Escalation path
AI Workflows
Multi-step work routed across models, tools and people without dropping state.
State, retries and idempotency are handled explicitly, so a failed step resumes instead of restarting the process. Somebody can read the run history six months later and understand what happened.
A failed step resumes from the last good state.
- State model
- Retry policy
- Run history
Software Systems
The service layer, data contracts and interfaces that make the system dependable.
Typed contracts at every boundary, integration tests that exercise the real path, and an operational runbook written on day one rather than after the first incident.
Contract violations surface in CI, not in production.
- Service layer
- Data contracts
- Runbook
Field note 04 · Logistics operations
The system runs during the shift that matters, not in a sandbox.
Exception handling was rebuilt around one workflow with a human review point above five thousand. The team kept the process they already trusted and lost the overnight queue.
Intake, ERP, dispatch
Above $5,000
9 weeks
Engagement
Quietly, in four phases.
Four steps, each producing an artefact you can review. No discovery theatre, no capability decks.
Map the operation
Two weeks, side by side with the people doing the work.
We trace the process as it actually runs, including the exceptions nobody documented and the workarounds somebody built in a spreadsheet three years ago.
- Artefact
Define the system boundary
Which steps become automated, which stay human, and where the handoff happens.
Written down and agreed before anything is built. This is the document that stops scope from quietly expanding in week six.
- Artefact
Engineer the workflow
Models, tools and services assembled behind one interface.
Evaluation and observability are wired in from the first commit, not retrofitted. You see the same traces we do, from the first week.
- Artefact
Run and improve it
Traces, cost per run and failure rates in plain view.
The system gets tuned against evidence rather than intuition, and the runbook stays current as the operation changes around it.
- Artefact
Readout
Measured on the operation, not the pilot.
Representative figures from a production deployment. Every number traces back to a run log.
4,120
Workflows in production
Peak 6,900 on month-end close.
38
Manual handling removed
Measured against the pre-engagement baseline.
99.2
Completed without human touch
The remaining 0.8% is escalation by design.
$0.014
Fully loaded inference cost
Including retries and evaluation passes.
Start
Bring us the process that keeps breaking.
Send the workflow, the volume and the failure mode. You will get a considered read on whether it should be engineered, and what that would take.
- Reply within two working days
- Written scope before any build
- No discovery call required to start