What you're building. A complete analytical engagement for the General Manager of a grower-owned specialty-berry cooperative in Salem, Oregon. The deliverables are four: a multi-method analytical notebook, a USDA grant-review brief, a cover letter to the client, and a handoff-ready engagement artefact another analyst could inherit.
What you're practicing. Decomposing a multi-faceted problem into distinct question types -- descriptive, inferential, causal, predictive -- and deploying each with the right method. Designing a validation architecture before any analysis runs. Two causal analyses (one quasi-experiment, one natural experiment) with refutation testing and substantive sensitivity. An inferential thread where a measurement change is the load-bearing covariate. A predictive screen that earns its place in the strategy. Synthesis that respects what each method can and cannot claim. An executive summary that leads on honest uncertainty for a regulator audience. AI infrastructure designed from scratch for an engagement that spans weeks.
What you're learning about the loop. This is the project where the structure transfers from the platform to you. No brief, no methodology memo, no colleague walking the analysis with you -- a senior colleague is reachable on demand for a sanity check, but she won't volunteer and won't hand you answers. You decide the question types, the preparation order, the validation per component, the synthesis logic, the audience design, and the AI infrastructure that supports all of it. AI is more capable here than at any point in the track -- and its defaults are correspondingly louder. It will reach for a single methodology when the problem needs five, treat a measurement change as a cleaning step rather than a covariate, miss the natural experiment unless you name it, bury caveats in the appendix. You catch each one. You are the methodological authority; AI is the execution layer.
Your Role
You are the senior data scientist hired for the capstone -- a multi-method engagement supporting a USDA Specialty Crop Block Grant review and the client's day-to-day decisions. The engagement runs across weeks. The deliverables go to a federal regulator and back to the client.
Every concept from the track -- description, prediction, forecasting, inference, causation, multi-method synthesis, inheritance review -- is back in scope, deployed selectively. The AI relationship is orchestration across phases: phase-scoped tool connectivity, per-agent context scoping, delegation decisions including the ones where you decide not to delegate. The directing repertoire you have built from P1 onward holds the line.
What's New
Last time you produced an inheritance review for a public transit authority: a verification map, a corrected analysis, and a plain-language brief for the ministry. The judgment work was salvage-vs-redo per finding, on someone else's notebook.
This time the seat inverts. You are not reviewing someone else's work -- you are producing work that would survive the same review you applied last time. The first contact is a voice memo with ambient sort-line sounds; the client names many problems and no specific questions. No colleague works the analysis alongside you -- earlier projects had a senior colleague at your elbow at the load-bearing moment; here one is reachable on demand for a sanity check, but the judgment stays yours and she won't volunteer. The engagement spans weeks, so context lifecycle management is part of the architecture.
The hard part is holding global coherence across five question types and across weeks of work, while the data has constraints that only surface if you ask the right questions in discovery. AI will not surface those on its own. You will.
Tools
Carry-forward. No new packages.
- Python 3.11+ in the conda "ds" environment, Jupyter, pandas, Polars, scikit-learn, statsmodels, DoWhy, matplotlib / seaborn, plotly
- MLflow -- a new
cascade-berry-capstoneexperiment with methodology-version tags - DuckDB, DuckDB MCP server, Jupyter MCP server, MLflow MCP server -- connected selectively by phase
- Claude Code with the assumption-checking skill (including the cross-agent composition check) and the pre-evaluation hook. You author one new engagement-specific skill, and optionally one new hook
- Codex CLI with AGENTS.md -- cross-tool review on the load-bearing causal specifications
- GitHub Copilot async delegation for the multi-specification sensitivity sweep, Git / GitHub
The data is new -- five years of multi-source records across roughly 800 farms, in US dollars (USD).
Materials
The project lives at ~/dev/data-science/p-18. The materials zip downloads from https://learntodriveai.dev/materials/data-science/p-18/materials.zip.
What is provided:
- Thin starter
CLAUDE.mdandAGENTS.mdfiles -- you author the engagement-specific content across the phases - A one-page cooperative context document (
cascade-berry-context.md) - A one-page USDA grant-review overview (
usda-grant-review-overview.md) -- what the regulator asks for at the programme level - A data dictionary (
data-dictionary.md) describing each of the six source files honestly - The raw multi-source data under
data/-- delivery records, pricing, facility logs, buyer contracts, member demographics, USDA grant beneficiary list
No question framings, no methodology memo, no analysis specification, no validation strategy, no deliverable templates, and no colleague producing any of it for you -- a senior colleague is reachable on demand for a sanity check, on-demand only and never volunteering. The methodology, validation architecture, analyses, synthesis, brief, cover letter, and handoff-ready engagement artefact are yours to produce.