Opportunity mapping
We sit with your team for two weeks, watch the work, and come back with a ranked list of what is worth automating and what is not.
Every service
Scoped as standalone pieces of work
Each one is scoped as a standalone piece of work. Most clients start with one and add the next once the numbers hold.
Model selection and tuning
Open weights, hosted APIs or a small task-specific model. We test candidates against your data before anyone signs a licence.
Workflow engineering
The unglamorous part: queues, retries, human review steps and the handoffs that decide whether anyone actually uses the thing.
Governance and controls
Audit trails, access rules, red-team runs and a written record of what the system may and may not decide on its own.
Evaluation and monitoring
A test suite built from your real edge cases, running on every change, with alerts when quality drifts instead of after a complaint.
Handover and training
Your engineers run it after we leave. Documentation, pairing sessions and a 90-day window where we stay reachable.
Delivery path
Four stages, in this order, every time
Order matters here — each stage produces the input the next one needs. We do not skip ahead, and we stop at any gate that fails.
Ground truth
We measure the current process by hand: volume, cost per item, error rate, where time goes. Nothing gets built against a guess.
Weeks 1–2Narrow prototype
One workflow, one team, real data. The goal is a decision, not a demo: is this worth the integration cost or not.
Weeks 3–6Production build
Hardening, permissions, monitoring and the review steps your risk team asked for. Rolled out to one site before all of them.
Weeks 7–14Handover
Your team takes the keys. We shadow for a month, answer questions for two more, then step back for good.
Weeks 15–18
4.2×
median return on the first workflow shipped
How we work
We stay until the numbers move, then we leave
Most AI work stalls somewhere between a convincing prototype and a Tuesday morning shift. Our whole practice is built around that gap.
-
One team, start to finish
The people who write the strategy write the code. No handover to a delivery pod you have never met.
-
Targets written into the contract
Cost per item, cycle time, accuracy floor. If we miss them, the last invoice adjusts.
-
Your data stays yours
Deployment inside your account or on your hardware. Nothing trains on your records without a signed instruction.
-
No lock-in by design
Standard tooling, documented interfaces, and a swap path for every model we recommend.
Sectors
Where our reference cases live
We take work outside these six, but slower — we will tell you if the learning curve is on your budget.
Banking and insurance
Claims triage, document review, adverse media checks and the audit trail your regulator will ask for.
31 engagementsHealth systems
Prior authorisation, coding support and clinician-facing summaries with a named reviewer on every output.
18 engagementsLogistics
Exception handling across freight documents, customs paperwork and carrier correspondence.
22 engagementsIndustrial
Maintenance history search, defect classification and shift-handover notes that survive the handover.
14 engagementsRetail and marketplaces
Catalogue enrichment, fraud review queues and support desks that resolve without a scripted loop.
19 engagementsPublic sector
Casework backlogs and correspondence handling, built for records requests from day one.
9 engagements
Questions
Asked before every kickoff
If yours is not here, write to us and we will answer it in plain language.
No. Waiting for clean data is the most common reason projects never begin. We work with what exists, document what is missing, and fix the records that actually affect the outcome.
Inside your cloud account, your data centre, or an isolated environment with no outbound network. We have delivered all three. The choice is yours and it does not change our fee.
The final invoice reduces on a scale written into the statement of work. We would rather agree that up front than argue about it at the end.
Only on your written instruction, only on your infrastructure, and only for a model you own. Vendor terms are reviewed and summarised for you before anything is signed.
We prefer it. Projects with two or three of your engineers embedded hand over faster and hold their results longer. It also reduces the fee.
Assessments usually begin within three weeks. Build engagements depend on team availability — as of this quarter, the next opening is in about six weeks.
Next step
Bring us the process nobody wants to touch
Send a short note about the workflow and the number you need to move. You will get a written reply from an engineer, not a form response, within two working days.