A Full Analytical Project: From a Symptom to a Delivered Decision.

Sokha Meas runs operations for a garment manufacturing group in Phnom Penh, Cambodia — three factories, 500 workers — and her margin has fallen from 12% to 7% over two years with no clear reason why.

You're practicing every discipline skill the track has built, together, in one project: framing the questions from a symptom, defining metrics so three structurally different factories can be compared fairly, designing the data acquisition strategy across five scattered systems, running the analysis, judging whether an experiment is warranted, building deliverables for several different audiences, and designing verification coverage across the whole chain.

What's different about directing AI this time is that the work has no provided shape and AI cannot supply one. AI will generate a plan that investigates everything equally and answers a question nobody asked. It will compare the three factories on raw output and confidently rank the wrong one first. It will produce one report when five audiences each need a different framing, and it will report what it found and stop — never stating what was inconclusive. AI executes every step well. The coherence across all of them — how the question framing constrains the metrics, how the metrics make the factory comparison fair or misleading, how the analysis feeds five deliverables without contradicting itself — is yours to hold. That is the whole job here.

Your Role

You're the analytical lead for the COO of a private garment group, owning the project end to end: from a one-line symptom to a delivered decision, a recommendation with projected impact, and a monitoring capability she keeps using after you're gone. She wants to know which fixes recover the most margin — and to keep watching once you're done.

What's different from last time is everything that used to be provided. There is no estate to evaluate, no reference framing, no runnable environment to inherit, no method to adapt, and no scripted help. You design and build all of it from a blank workspace. The relationship with AI is fully autonomous: you set the decomposition, the context strategy, and the connectivity yourself, and you decide what stays your direct judgment — metric governance, question framing, and how you talk to the client are not delegated.

What's New

Last time you evaluated a working analytical estate against a forced platform migration — the dashboards, the metric layer, the infrastructure, and a target environment were all built; you judged what survived the move and rebuilt it with equivalence proven number by number.

This time nothing is built. You receive a symptom and raw, unjoined data — and you author the entire project: the questions, the governance, the acquisition strategy, the analysis, the experiment-or-not decision, the deliverables, the infrastructure, and the verification memo. No skeleton, no template, no scripted colleague stepping in at the right moment.

The hard part is holding the whole analytical narrative coherent across a multi-week project when AI will only ever see one piece at a time. The instinct will be to analyze everything the data offers. The harder, more valuable move is deciding what to ask, what to leave directional, and how to keep one definition meaning the same thing from the first metric to the final board slide.

Tools

Everything here is familiar — nothing new is introduced. Claude Code as the directing environment, DuckDB as the analytical engine, Python for statistical work, dbt as the metric governance layer Sokha keeps, Codex CLI for cross-AI verification, Git and GitHub with conventional commits, and MCP connectivity you design across the source systems. The deliverable format — a BI dashboard, a Streamlit app, or something else per audience — is your call. What's new is that you select, connect, and infrastructure all of it for a project of your own design.

Materials

• The first-contact email — Sokha's symptom, her own hypothesis, and a request to discuss scope, in the project chat. The discovery is the thread, not the email.
• The raw factory data under factory-data/ — five separate, unjoined source exports, exactly as they come off five systems that have never been unified: factory-data/production/, factory-data/quality/, factory-data/payroll/, factory-data/procurement/, and factory-data/orders/. The three factories' files are in different shapes because they are different systems.
• CLAUDE.md — pure project context: the client, the business, the symptom, the source systems, and the standing instruction that metric governance, question framing, and stakeholder communication stay your direct judgment. No questions, no metric definitions, no data dictionary, no work breakdown — you author all of that.
• The senior colleague channel — Raj Patel, available on demand for capstone consultation. Not scripted into the work.

Download the materials zip (https://learntodriveai.dev/materials/analytics/p-23/materials.zip) and unzip it to ~/dev/analytics/p-23.