You are running a donor-review preparation engagement for the executive director of a Senegalese microfinance nonprofit. The deliverable is a complete communication package built from three distinct analyses — descriptive portfolio health, inferential default drivers, and predictive at-risk loans — synthesized into one coherent picture a donor can read and act on.
You profile five years of loan, repayment, borrower, and regional-economic data; surface three load-bearing constraints (a recent shift from centralized to delegated loan approval, a seasonality interaction between agricultural and year-round vendor lending, a regional office whose high default rate reflects serving the most vulnerable borrowers); decompose the request across the five question-type categories; design the preparation and validation architectures before any analysis runs; run the three methods in parallel; consult Dr. Nadia Petrova on the synthesis; and produce the package itself.
The AI-direction lesson is that the communication package IS the deliverable, not the notebook. AI defaults to a single composite analysis when the engagement needs three. AI buries limitations in fine print. AI generates static charts when the donor needs interactive ones. Your value is the synthesis discipline — keeping the three methods coherent, moving uncertainty to the front of the executive summary, and writing a recommendation that respects what each method can and cannot claim.
Your Role
You are the senior data scientist on the engagement, hired for the donor-review preparation. You design the engagement frame from the meeting transcript, surface the three hidden constraints through Living Client conversation, design the preparation and validation architectures before any analysis runs, run the methods in parallel, synthesize the results, and write the package the donor will read.
The familiar terrain is what you've built across the prediction projects — leakage-free preparation, assumption checks, sensitivity, baseline comparison, multi-agent coherence. What is new is communication ownership at full strength. You design the layered package shape, decide where uncertainty lives in the prose, and direct AI through interactive visualizations built for a non-technical donor. The artifact is also designed for inheritance — another analyst should be able to pick up your repo and extend the work.
What's New
Last project, you produced a council-facing recommendation for a Canadian municipal water utility across three parallel tracks. The hard part was holding "do not build it" as a defensible professional output.
This project shifts the synthesis problem. The three parallel methods are not three approaches to one question — they are three different question types running together as the deliverable's evidence base. Description, inference, and prediction each address a different component, each with its own validation, woven into one picture. The communication package replaces the single report: executive summary with uncertainty in the same paragraph as the headline, detailed findings per method, interactive visualizations as the donor's drill-in path, conditions for change, action items. Plotly enters as a new technical layer; Dr. Nadia Petrova appears once on the synthesis.
The hard part is the synthesis itself. AI produces three method summaries that sit beside each other; weaving them into one coherent recommendation is your work. The register also shifts — warm, mission-driven, French business expressions in formal English, walls of text where context matters, borrower-story tangents that carry analytical weight if you are listening.
Tools
Carry-forward, with one new technical layer.
- Python 3.11+ in the conda "ds" environment, Jupyter, pandas, scikit-learn, statsmodels, matplotlib / seaborn
- MLflow — a new experiment on the existing local file store
- DuckDB, DuckDB MCP server, Jupyter MCP server, MLflow MCP server
- Claude Code with the assumption-checking skill and pre-evaluation hook active (you do not modify these)
- Codex CLI for cross-model review of the synthesis
- AGENTS.md, Git / GitHub
- plotly (NEW) — interactive visualizations are part of the deliverable. Installed into the project venv as a normal package.
The data is new — five years of microfinance records across six regions of Senegal, in CFA Franc (XOF).
Materials
The project lives at ~/dev/data-science/p-16. The materials zip downloads from https://learntodriveai.dev/materials/data-science/p-16/materials.zip.
What is provided:
- A project memory file framing the engagement and the Teranga Microfinance domain
- Data across six regions and five years (~30,000 loans): records, repayments, borrower demographics, loan officer assignments, regional economic indicators
- A data dictionary, honest about field semantics but not volunteering the three hidden constraints — those surface through Living Client conversation
- A multi-method decomposition template, a preparation architecture template, a validation architecture template, a communication-package skeleton, and the cross-review focus template
No baseline analysis, no per-method specifications, no draft executive summary, no recommended approach. The decomposition, the preparation and validation architectures, the three parallel analyses, the synthesis, and the package itself are yours to design.