Applied data science with revenue impact.
You need a model. Your team has the data and the ambition but not the bench depth. Past attempts produced something that lived in a notebook and never made it to production — or shipped to production and quietly broke six months later. The business question is clear; the path from raw data to a decision your team can act on is not.
- 01 Scoped model
Defined target, success metric, decision surface, and validation strategy before a single line of code. We don't build until we know what 'working' looks like.
- 02 Validated against holdouts
Out-of-time and out-of-sample validation. No leakage. We test against the data the model wouldn't have seen at decision time.
- 03 Production deployment
Model lives where it can be called — batch warehouse, real-time service, or embedded in your existing app. Your engineering partner approves the choice.
- 04 Monitoring + retraining
Drift detection wired in. A cadence for re-training and a runbook for when performance degrades. Six months from now you'll know whether to retrain or rebuild.
- 05 Handoff documentation
Every assumption, every feature definition, every model decision written down. Your team can rebuild it from the docs if they need to.
- 06 Decision integration
The model output lands in a place your team actually acts on — a dashboard, an automation, a CRM field. Models that nobody opens are just expensive notebooks.
- Teams with documented data and a single decision the model serves
- An engineering partner who can host model artefacts (in-house or vendor)
- A clear business owner — someone whose KPI moves if the model is right
- Companies past PMF with enough data history for meaningful validation
- Vague 'do something with AI' briefs
- Pre-data teams (we can't model what doesn't exist yet)
- Engagements without a named business owner
- Procurement-led shopping where the scope is decided by a vendor template
What's a typical engagement length?
8–16 weeks. Shorter than that means we cut corners on validation; longer means we're scope-creeping. We re-scope at week 4 if the data tells us a different problem.
Will you work in our stack?
Yes. Python or R on top of BigQuery, Snowflake, Databricks, Postgres — whichever you already have. I don't sell you a new platform.
What if the model doesn't beat the baseline?
That's a real outcome and you still get the answer: this problem isn't model-shaped. The engagement scope includes the null finding being delivered with the same rigour as a positive one.
Who maintains the model after you leave?
Your team. The handoff includes a walkthrough, a runbook, and a documented retraining trigger. About a third of clients extend into a quarterly check-in.
Will you sign an NDA / DPA?
Yes. Standard terms, fast turnaround. Larger engagements include a data-handling addendum we negotiate up front.
Bring the question. I'll bring the rigour. Engagements scoped on a discovery call — 20 minutes, no deck.