We are a decision intelligence consultancy, not a software vendor and not a staffing shop. Every service below ends with a model, a report or a routine that a named person uses to make a specific call — plus the measurement plan that proves whether it worked.
Most firms sell time. We sell a decision that gets better, so the shape of the work follows the size of the question. Three formats, each with a written hypothesis at the start and a measurement plan at the end.
Two weeks with your data, your P&L and the eight people who actually make the call. You leave with a written view of where better forecasting changes an outcome — and where it plainly would not.
Production pipelines and models, shipped in fortnightly increments against an accuracy bar agreed before anyone writes code.
For organisations building a durable capability rather than one model. We sit inside your team, ship alongside your engineers, and agree the exit plan in month one so the end is never a surprise.
Demand, revenue, risk and capacity models that hold up at horizon. We build to the horizon your decision actually needs rather than the one the tooling defaults to, and we report error where it costs money instead of on average.
Read from 01-demand-forecasting/run.py when this page was built — 738 lines. What that project imports and calls, not a list of everything we know.
Read from 02-data-governance/run.py when this page was built — 1,148 lines. What that project imports and calls, not a list of everything we know.
Operating model, architecture roadmap and the governance to keep it honest. We write the target state, cost it, sequence it and name the person accountable for each domain — before anyone signs a licence.
Warehouse, transformation and semantic layer built for change rather than for a demo. Tested, documented, version-controlled and boring in the best sense: an estate a new engineer can safely change in their first week.
Read from 03-analytics-engineering/run.py when this page was built — 1,178 lines. What that project imports and calls, not a list of everything we know.
pandas only — 18 names.
Read from 04-decision-intelligence/run.py when this page was built — 826 lines. What that project imports and calls, not a list of everything we know.
KPI trees, board-grade reporting and the rituals that make numbers get used. We start in the meeting where the decision is actually made and work backwards to the metric, then delete whatever nobody opens.
Models in production with monitoring, retraining and a clear rollback path. We ship fewer models than we are asked to, and we retire the ones that stop paying for themselves — the kill criterion is written before the modelling starts.
Read from 05-ml-production/run.py when this page was built — 1,015 lines. What that project imports and calls, not a list of everything we know.
Read from 06-pricing-elasticity/run.py when this page was built — 828 lines. What that project imports and calls, not a list of everything we know.
Elasticity, promotion effectiveness and margin recovery you can bank. We measure what your promotions actually caused, then hand commercial teams guardrails they can price inside without asking permission every time.
Four things leave with you on every engagement, whatever the format and whatever the service. None of them is a slide deck and a goodbye.
Twenty to forty pages of argument. What we found, what we recommend, what we would not do in your position, and what it costs you to be wrong about each one.
Your repo, your cloud account, your credentials. Pipelines, models, tests and CI, documented well enough that a new engineer ships a change in their first week.
Alert thresholds, on-call ownership, retraining triggers and the rollback path. Written for the Tuesday morning when something breaks and none of us are in the building.
Training for the people who use the number, a decision log for the meeting where it gets used, and a measurement plan that survives our departure by at least a year.
Fee models differ because risk differs. Outcome-linked means part of our fee sits against the accuracy bar we agreed in writing before the first sprint.
| Format | Typical duration | Fee model | Ends with |
|---|---|---|---|
| Decision Audit | 2 weeks | Fixed fee, agreed up front | A written diagnostic and a ranked opportunity list |
| Build Sprint | 8–16 weeks | Outcome-linked to the accuracy bar | Production models running in your environment |
| Embedded Practice | 6–12 months | Retained monthly, reviewed quarterly | Your team running the whole thing without us |
You do, outright, from the first commit. Work happens in your repository and your cloud account wherever possible, and the intellectual property transfers to you as it is created rather than on final payment.
We keep no copy of your data after an engagement closes, and the deletion is confirmed in writing. We reuse our own generic tooling and methods across clients; we never reuse your data, your models or your feature definitions.
No. Predictify Analytics has taken no reseller commissions and holds no technology partnerships. That independence is the whole reason a recommendation from us is worth reading.
In practice we work across the major warehouses, orchestration tools and BI layers, and we will happily tell you that the platform you already own is good enough — which is the answer more often than vendors would like.
No, and waiting for one is usually the more expensive choice. A Decision Audit needs interviews and a sample of data, not a finished platform.
What we will not do is build a forecast on a foundation that cannot be refreshed on a schedule. If the data cannot arrive reliably, fixing that is the first piece of work, and we will say so before you sign anything.
We do not guarantee accuracy, and you should be wary of a firm that does. What we agree instead is an accuracy bar: a specific metric, measured on a specific slice of your data, against your current method as the baseline. It is written down before the first sprint.
Every release is scored against that bar in a backtesting harness you keep. The number that matters is the one measured on your data, in your environment, against what you do today.
Alongside them, in the same repository, in the same stand-up. We take the modelling work that needs a specialist and a clear runway, and we hand it back documented.
Your engineers review our pull requests, not the other way round. If a member of your team wants to learn the modelling side, we pair with them deliberately — that transfer is written into the plan, not left to goodwill.
Handover is a fortnight of its own, not a final email. Your team runs the pipeline and the retraining cycle with us watching, we fix what breaks, and the runbook is corrected wherever reality disagreed with it.
Ninety days later we return for a measurement review: did the decision improve, and did the value show up in the P&L. It is a fixed part of the fee and it is not optional.
Send us the decision you are trying to improve — not the tool you are considering. We will tell you in writing which format fits and what it would cost, within one business day, and we will say so plainly if the answer is none of them.