Most advisory work is fairly ordinary, and we are happy to do it well. Occasionally a problem arrives that fits no category at all.
A typical engagement starts with feasibility. We spend around two weeks inside your data and systems, talking to the people who run them, and come back with a plan that says what is worth doing, in what order, and what it would take. The plan is written so that a board can act on it. If it holds up, we build the proof of concept, together with the evaluation that shows whether it actually works. Sometimes the job is simpler and amounts to getting a capable model running inside your firewall, and getting your staff to genuinely use it.
The scope and fee are fixed in writing before we begin, and everything we produce transfers to you at the end: code, documentation, models. The fee moves with the scope in both directions. If we find work that wasn’t in the plan, we tell you what it would cost before doing any of it; if part of the plan turns out to be unnecessary, we say so and take it off the invoice.
Then there is the other kind. Someone suspects a pattern in the data but cannot prove it. Ten years of inspection reports have never been read end to end. One factory’s energy bill runs a third above its twin’s, and the two make the same products on the same machines. A system fails intermittently in a way nobody can reproduce, or a question matters a great deal but cannot quite be written down as a tender document. This is investigative work, where the evidence has to be assembled and reasoned from before anything can be concluded.
These engagements are short and the brief is narrow. A question goes in, and what comes out is the answer with the workings shown. The tool is whatever the question needs, and often that is a join, a distribution and a week of careful reading rather than a model.
We’re a new firm. The cases below are invented. They show the kind of work we mean.
A manufacturer’s ERP exports reports in a format that only one long-serving employee can still interpret. An agent is set up to parse them, which works, but is expensive to run. An overseer watches the agent, identifies which of its corrections are routine, and replaces them with ordinary deterministic code, supplier by supplier. After a few months the model is only handling genuine exceptions, and the ordinary code runs the rest.
Exemplum · hypothetical
We have since built this twice on engineered data, per report type in the legacy ERP case and supplier by supplier in the invoice case.
The monthly board pack shows margin slipping, and the variance analysis attributes it to material costs. An agent sent through the underlying ledger finds a different explanation. A single vendor changed its invoicing format in March, and freight charges have been quietly reclassified ever since. The headline attribution was not wrong, but the useful number sat three levels below the one being reported.
Exemplum · hypothetical
We have since built this on engineered data, as three blinded agent runs through the ledger behind a board pack in the board-reporting case.
Ten years of inspection reports sit in PDF, filed and unread. An agent reads all of them and returns three patterns worth human attention, including one that predicts a class of equipment failures previously assumed to be random.
Exemplum · hypothetical
You deal with the two of us. The first conversation establishes whether we’re the right people, and if not we say so and suggest alternatives. The work is done by whoever you spoke to, the fee doesn’t move once agreed, and everything ends with documentation and a proper handover.
Tell us the problem, what is at stake, and when you need to move.