datapot-fabric-agent-demo

Power BI analytics-cycle demo, built with AI agents — as reviewable code, not clicks.

A Microsoft Fabric / Power BI monorepo where semantic models and reports are authored with AI agents (Claude Code and GitHub Copilot) in diff-able PBIP format (TMDL + PBIR) — then validated, documented, and version-controlled like software. Report 01 is the first demonstration scenario; the core goal is to show reusable agentic patterns across the full analytics cycle.

📖 See exactly how it was built — prompts, debug log, and cost: read the build recap (rendered via GitHub Pages; source in docs/session-recap.html).

🔬 Research & educational use only · 100% dummy data. Experimental; not production-tier and not through Datapot's QA. The bundled "BankDIAD" dataset is entirely AI-generated — no real people, customers, or institution. Don't use the model or numbers for real decisions. No warranty. Full notice ↓

---

What it produces

A clean, conventional star schema — 6 dimensions, 3 facts, and a disconnected measures table:

The report has three pages — Overview, Channel Performance, and Branch Scorecard (31 visuals). Open BranchChannelPerformance.pbip to explore them (Getting Started ↓), or skim the build recap for the visuals and story.

Project intent: demo AI-agent capability across the analytics cycle

This repo is designed as a reference demo for each analytics stage, not as a deeply specialized banking implementation:

1. Discovery & scoping
2. Data contract & profiling
3. Semantic modeling (TMDL)
4. Report authoring (PBIR)
5. Validation & review
6. Publish / operate in Fabric

You can reuse the same approach for other domains by creating new report folders under reports/.

How report development workflows work in this repo

Every report follows the same lifecycle under reports/NN-name/:

1. Scope the business goals in docs/report-spec.md.
2. Define required input files/columns in dataset/DATASET-CONTRACT.md.
3. Add data to dataset/raw/, then profile it (docs/data-profile.md).
4. Build/adjust the semantic model in TMDL (pbip/<Project>.SemanticModel).
5. Build pages/visuals in PBIR (pbip/<Project>.Report).
6. Keep dictionary and build log updated (dictionary/, docs/build-log.md).
7. Run validation gates before commit (tmdl-validate, validate_pbip.py, jq empty).

For full details, see docs/report-lifecycle.md and CONVENTIONS.md.

Workflow examples

• Existing example: reports/01-branch-channel-performance/ shows a complete end-to-end implementation (dataset contract → model design → report spec → TMDL/PBIR artifacts).
• New report example: create reports/02-customer-retention/ using the same contract, then deliver:
  1. KPI scope (Retention Rate, Churn Rate, Active Customers),
  2. a star schema (Date/Customer/Segment/Region + facts),
  3. three pages (Overview, Drivers, Scorecard),
  4. validation and registry update in reports/README.md.

What you'll learn

A worked, end-to-end example of agentic analytics engineering on the Microsoft data platform:

• 🧱 PBIP as code — how a Power BI model + report look as reviewable TMDL/PBIR files, not a binary .pbix.
• ⭐ Star-schema modeling — 6 dimensions, 3 facts, hidden surrogate keys, explicit measures only.
• 🧮 DAX measure design — 22 measures in display folders (volume · revenue & targets · service/NPS · digital adoption · time intelligence).
• 📒 Spec-driven, documented builds — dataset contract → data profile → model design → report spec → build log → data dictionary.
• 🤖 Agent-fleet workflow — how to route tasks to the right specialist skills in Claude Code and GitHub Copilot.
• 🧰 Reusable conventions — folder, naming, TMDL & PBIR rules that keep a multi-report repo consistent (CONVENTIONS.md).

Agent fleet (recommended routing)

Need	Best fit
Cross-workload Fabric implementation (lakehouse/warehouse/pipeline/notebook)	FabricDataEngineer
Power BI business Q&A from reports/models	FabricIQ
Fabric governance/ops/security/capacity tasks	FabricAdmin
App integration with Fabric APIs/data	FabricAppDev
Local PBIP/TMDL/PBIR build and validation	power-bi-agentic-development skills

Quick start for non-technical, business-focused analysts

Start here for plain-language, low-code workflows in both tools:

• Business analyst quickstart (Claude Code + GitHub Copilot)
• Claude walkthrough video: How to use Claude Code (YouTube)

Keep skills-for-fabric current

For GitHub Copilot CLI plugin users:

/plugin update fabric-skills@fabric-collection

If you still use the legacy id, this alias also works:

/plugin update skills-for-fabric@fabric-collection

Getting Started — open & explore Report 01

Prerequisites

• Power BI Desktop (latest; free from the Microsoft Store).
• The PBIR preview features enabled (one-time, below) — required, or the report won't open.
• (Optional) Python 3 + openpyxl — only to run validate_pbip.py or regenerate the dummy data; not needed to open/refresh the report.

1. Clone

git clone https://github.com/DatapotAnalytics/datapot-fabric-agent-demo.git
cd datapot-fabric-agent-demo

2. Enable the PBIR preview (one-time, required) — in Power BI Desktop: File ▸ Options and settings ▸ Options ▸ Preview features → tick Power BI Project (.pbip) save option and Store reports using enhanced metadata format (PBIR) → OK → restart Desktop. (Skip this and Desktop may refuse to open the project or show "The report has no pages".)

3. Open the project — reports/01-branch-channel-performance/pbip/BranchChannelPerformance.pbip

4. Point the model at your clone (required) — the data path is parameterized. On the ribbon: Home ▸ Transform data ▸ Edit parameters, set DataFolder to the absolute path of the raw-data folder in your clone, e.g.

C:\Users\<you>\src\datapot-fabric-agent-demo\reports\01-branch-channel-performance\dataset\raw

5. Refresh — Home ▸ Refresh. The dummy data loads (36 monthly partitions, 2023-01 → 2025-12). It's 100% synthetic, so open and refresh with no privacy concerns.

The three pages

1. Overview — headline volume & value: KPI cards (Total Transactions, Gross Transaction Value, NPS, Digital Adoption %), a monthly trend, a by-channel breakdown, and a Year slicer. How much is flowing, and how is it trending?
2. Channel Performance — the digital shift: digital-adoption KPIs, a 100% stacked column of channel mix over time, volume & average service time by channel, and a Channel × Transaction-Category matrix. How much activity is digital, and is it growing?
3. Branch Scorecard — per-branch performance: a Region › City › Branch scorecard matrix, NPS by branch, and a detractor-reason breakdown. Which branches lag on service, NPS, and targets?

Repository layout

datapot-fabric-agent-demo/
├── README.md                  ← you are here
├── CLAUDE.md                  ← how an agent (or human) should work in this repo
├── CONVENTIONS.md             ← naming, folder, PBIP/TMDL & dictionary standards
├── CONTRIBUTING.md · SECURITY.md · CODE_OF_CONDUCT.md · CHANGELOG.md · LICENSE
├── docs/                      ← repo-wide docs
│   ├── report-lifecycle.md    ← design → dataset → model → report → deploy
│   ├── business-analyst-quickstart.md ← non-technical starter guide for Claude/Copilot
│   ├── business-glossary.md   ← banking terms shared across reports
│   ├── fabric-environment.md  ← target workspaces / deployment (fill in per tenant)
│   ├── assets/                ← diagrams (star-schema.svg)
│   └── session-recap.html     ← how this build actually went (prompts, debug log, cost)
├── shared/                    ← assets reused by every report
│   ├── themes/                ← the Datapot brand Power BI theme
│   └── templates/             ← starter files for a new report
└── reports/                   ← ALL reports live here, one folder each
    ├── README.md              ← report registry / index
    └── 01-branch-channel-performance/
        ├── dataset/  dictionary/  docs/  pbip/   (+ README.md)

The per-report contract

Each report folder always carries these four things — nothing leaks between reports:

Component	Folder	What it is
Folder	reports/NN-name/	One numbered folder per report
Dataset	dataset/	Source data + DATASET-CONTRACT.md describing the expected files/columns
Document	docs/	report-spec.md, model-design.md, build-log.md, data-profile.md
Dictionary	dictionary/	data-dictionary.md + .csv — every table, column, and measure defined

Power BI artifacts (semantic model + report) live in pbip/ as source-controllable PBIP (TMDL semantic model + PBIR report).

Reports

#	Report	Domain	Status
01	Branch & Channel Performance	Banking	🟢 Complete — 3 pages; opens & refreshes in Power BI Desktop
02	(planned)	TBD	🔴 Ideas welcome — open an issue

New reports are added as self-contained folders over time — ⭐ star / watch to follow along. Legend: 🔴 planned · 🟡 in progress · 🟢 built. Full registry: reports/README.md.

Adding a new report

See CONVENTIONS.md and docs/report-lifecycle.md: copy shared/templates/ into a new reports/NN-name/, write the dataset contract, paste & profile data, build the TMDL model, then the PBIR report, validate, and (optionally) publish to Fabric. Contributions welcome — start with CONTRIBUTING.md.

Local development & validation (contributors)

tmdl-validate <file>.tmdl                                   # TMDL structural lint
python validate_pbip.py reports/NN-name/pbip --no-pbir-cli  # whole-project check (0 errors)
jq empty <file>.json                                        # PBIR JSON syntax

Windows / Claude Code note: if your clone path contains non-ASCII characters or spaces, the pbip plugin's PostToolUse PBIR hook may mis-read it and emit false "missing field" errors on .pbir / .Report writes — the files are fine if jq empty and validate_pbip.py pass. See CLAUDE.md for the full TMDL/PBIR rule set (folder naming, LF/no-BOM, theme metadata).

About Datapot

Datapot Analytics is a Vietnam-based data-analytics training and consulting company. We help people and teams become fluent in modern analytics — SQL, Power BI, data modelling, and the Microsoft data platform. This repo is part of our applied research into agentic, AI-assisted analytics engineering, shared openly so the community can learn the workflow — what works, what breaks, and why.

Explore & contribute

• ⭐ Star this repo if the agentic-PBIP workflow is useful — it helps others find it.
• 🔎 Explore Report 01 → Branch & Channel Performance
• 📖 Read the build story → session recap
• 🎓 Learn analytics with us → datapot.edu.vn
• 💬 Bug or idea? → open an issue · see CONTRIBUTING.md

License

Code is released under the MIT License. The bundled BankDIAD dataset is 100% synthetic and free to reuse for learning. Documentation and report content © Datapot Analytics — reuse encouraged with attribution.

🔬 Data & status notice

This repository is an experimental, research-purpose project. It is not production-tier and has not passed Datapot's own production standards / QA workflow. Structure, models, conventions, and data may change at any time without notice. Do not rely on this repository, its model, or its numbers for real decisions. No warranty; use at your own risk.

Data: the bundled "BankDIAD" dataset is 100% dummy, AI-generated data — every name, date of birth, balance, branch, and value is fabricated for demonstration. It contains no real, private, or personal data, represents no real customers or institution, and violates no one's privacy.