Most models get cut. We ship the one that runs.
We build forecasting, routing and classification systems on your operational data, then run them in production behind the tools your team already use. Deployment is the deliverable, not the demo.
02
03
04
Six approaches. One survives contact with your data.
01The cut
The hard part
The hard part was never the model.
A forecast that is four percent more accurate is worth nothing until something downstream changes because of it.
Most of the work in an AI project is not the model. It is the plumbing around it: getting to the data, agreeing what a row means, deciding who is accountable when the system is wrong, and wiring the output into the tool where someone actually makes the decision.
That work is unglamorous, and it is where projects die. Teams ship a convincing prototype, present it, then discover the warehouse table it depended on is rebuilt nightly by a script nobody owns. We start from the other end: the decision first, then backwards to the data, and we build only what the decision needs.
We would rather ship a blunt model that runs every day than a precise one that runs once.
02The bench
6–12 weeks
Three phases. Six to twelve weeks.
Every engagement runs the same strip. Some stop at the first frame, and that is a good outcome.
01Instrument
We map one decision end to end — who makes it, how often, on what evidence, and what it costs when it is wrong. Then we find out whether the data to support it actually exists.
Weeks 1–2Marked in
02Model
We build the smallest system that beats the current process, measured against the decision itself rather than a held-out test set. Gradient-boosted trees more often than transformers. We tell you which and why.
Weeks 3–7On the bench
03Operate
The system goes behind your existing interface — the CRM, the dispatch board, the ledger — with monitoring, versioned decisions and a rollback path. Your team runs it. We hand over the code and the runbook.
Weeks 8–12Taped — in production
03The panel
eu-west-1
Every system you run, and what it is doing right now.
Each deployed system reports its own health. Input drift, decision volume, latency at the ninety-fifth percentile, and the model version behind each decision. State is a mark, not a colour — a drifting system gets struck, not tinted.
- Taped — running in production
- Struck — drifting, pulled for rework
- On a pin, deferred
- On a pin — built, not yet committed
smzone / production
eu-west-1
| System | Drift | p95 | Decisions / 30d | Mark |
|---|---|---|---|---|
| Demand forecastfx-demand-04 | 0.02 | 31ms | 1.24M | Taped, in production |
| Route assignmentrt-dispatch-11 | 0.05 | 84ms | 402K | Taped, in production |
| Ticket triagecl-intake-02 | 0.31 | 52ms | 918K | |
| Ledger anomaliesan-ledger-07 | 0.01 | 19ms | 77K | Taped, in production |
| Contract extractionex-contract-01 | — | — | — | On a pin, deferred |
04The strip
Drag or scroll
What we build.
Demand forecasting
Stock, staffing and capacity at the granularity you actually order at — per SKU, per site, per day.
Routing & assignment
Constraint-aware allocation of jobs, vehicles or cases, re-solved as conditions change through the day.
Classification & triage
Intake sorted, tagged and prioritised on arrival, with confidence thresholds that escalate to a human.
Anomaly detection
Ledger, telemetry and process monitoring that flags the unusual rather than the merely rare.
Document extraction
Structured fields pulled from contracts, invoices and forms, with per-field confidence and an audit trail.
Retrieval systems
Question answering grounded in your own documents, with citations and a hard refusal when the evidence is absent.
05The bin
Declined
What we hang on a pin.
Nothing is deleted at a bench — it hangs where you can reach it. These are the jobs we routinely set aside, and why.
A chatbot on your website
Rarely tied to a decision that costs money. If it is, say which one and we will look again.
A model with no owner
If nobody on your side is accountable for the output, it will not survive its first bad week.
Anything built on a table nobody owns
We will find this in week one and tell you before you spend the rest of the budget.
Replacing a team
We build systems that make a decision faster or more consistent. Headcount is your call, not our pitch.
06Note
What a client said
They spent the first two weeks telling us our data could not support the thing we asked for, then built the thing it could support. It has run every morning since March.
Tell us about one decision.
Not a roadmap. One decision your team makes often, and what it costs when it goes wrong.
Start a pilot