Design the study so it answers your question
Most analysis fails before any data is collected, because the study was never pointed at the decision. We start with the decision: what are you choosing between, what would change your mind, and when do you need it? That becomes a scoping memo, and the scoping memo becomes the design.
Estimate the model so the numbers can be trusted
What people say they will do and what they actually do are different things, and the gap between them is where money is won and lost. Where the question allows it, we measure behavior rather than opinion — incentivized experiments, discrete-choice designs, revealed preference. Where it does not, we say so and we quantify what that costs you in confidence.
Every engagement carries an assumptions register: every input, its source and its plausible range. You should be able to disagree with an assumption, change it, and see what happens to the answer.
Deliver the tool your team keeps using
Some engagements end in a decision, some in a tool your team uses every day, and most in both. Pipelines, dashboards, models, maps, scripts and documentation go to you at the end, in formats you can open and rerun.
Use AI where it helps, and check it
Language models save real time on reading, sorting and drafting, and we use them for that. Nothing a model produces reaches you unchecked: we build verification into the pipeline, test models against your own material before choosing one, and keep confidential data on local models that never send it to an outside service.
Our tooling policy: open by default
We are comfortable in most proprietary software in our field — ArcGIS, Stata, Qualtrics, the Microsoft suite — and we will work in your licenses when you already own them and want the result there. But wherever we can, we build on open source, so you never have to buy a license to run, audit or extend what we deliver.
| Need | Open-source default | Also work in |
|---|---|---|
| Statistics and modeling | R; Python (statsmodels, PyMC, Stan) | Stata, Mathematica |
| Choice experiments and survey fielding | oTree, LimeSurvey | Qualtrics |
| Maps and spatial analysis | QGIS, PostGIS, Leaflet | ArcGIS Pro and Online |
| Dashboards and apps | R Shiny, Dash, Streamlit | Power BI, Tableau |
| Language models | Open-weight models run locally (llama.cpp, Ollama) | Hosted models under your own agreements |
| Documents and workbooks | LibreOffice, open formats | Microsoft Office |
| Data storage | PostgreSQL, SQLite, CSV, Parquet | SQL Server, Excel |
Deliverables come in open formats that still open cleanly in ArcGIS, Excel, Stata and Power BI. Where a project needs a licensed data source, we say so up front and pass it through at cost — and where the decision allows, we use a lower-cost public source so you can rerun the model without buying anything.
What you can expect
- A scoping memo before any work is billed
- An assumptions register you can argue with
- A written recommendation with a range, not a single number
- Code and data handed over, documented, and yours
- Direct access to the people doing the work, because there is nobody else to pass you to