Data Intelligence Lab
Reproducible demos across data governance, customer analytics, privacy and governed AI.

---

What this is

A single home for the demos and reference implementations behind four areas of work: data governance, customer analytics, data privacy and governed AI.

Each track holds self-contained projects. They are evidence, not products — every project is labelled with what it actually is, so nobody mistakes a 2025 experiment for something under active maintenance.

Tracks

Track	What lives here
data-governance/	Catalog-aware agents: reading metadata, and using it as context before touching data
governed-ai/	LLM agents with grounded access to data — SQL, RAG, classification
customer-analytics/	Segmentation, churn, next-best-offer, ARPU, incrementality — on one synthetic, causal, reproducible telco data model
data-privacy/	(empty) consent, minimisation, subject rights

Projects

Project	Track	What it does	Maturity	History
openmetadata-mcp-agent	data-governance	Conversational layer over an OpenMetadata catalog — ask in natural language, answered via MCP tools	reference	35 commits · 2026-03 → 2026-06
agent-data-analyst	data-governance	Analyses the data using the catalog as governance context: who owns the table, did it pass quality, what the codes mean	reference · paused	47 commits · 2026-02 → 2026-06
agent-sql-khipu_ai	governed-ai	LLM agent with SQL access and a RAG knowledge base, local models	archived snapshot	8 commits · 2025-01 → 2025-11
agent_text_classification	governed-ai	Sentiment classification with a local LLM	archived snapshot	8 commits · 2025-02
telco-customer-intelligence	customer-analytics	Synthetic, causal, reproducible telco data model + the analytics cases it feeds (segmentation, churn, NBO, ARPU, incrementality)	work in progress	data model + churn and incrementality cases built & tested; 3 cases to go

The two agents in data-governance/ are a progression, not a duplicate. openmetadata-mcp-agent talks to the catalog. agent-data-analyst analyses the data, using the catalog as context first. Reading them side by side is the point.

Maturity labels

Label	Means
maintained	Actively developed, issues get answered
reference	Complete and working; kept as a reference implementation, not actively extended
demo / POC	Illustrates one idea; not built to be deployed
archived snapshot	Preserved as it was written, not modernised

archived snapshot is literal: those directories are byte-identical to the repos they came from. Diff them against the originals — nothing was retouched. Their linter findings are silenced in pyproject.toml rather than patched, because editing a snapshot to look like it was written today destroys the only thing it is good for.

History

Every project keeps its full original history. Migration used git filter-repo --to-subdirectory-filter, so commits already point at the track path — git log, git log --follow and git blame all work from the new location:

git log --oneline -- tracks/data-governance/agent-data-analyst
git log --follow -- tracks/data-governance/agent-data-analyst/agent.py

(A plain git subtree add leaves the commits pointing at the old paths: the history is in the graph but is not reachable from the track. That is the difference between moving files and preserving history.)

Running a project

Each project keeps its own README, .env.example and dependencies. Start there:

cd tracks/<track>/<project>
cat README.md

Nothing here needs credentials to read. Projects that connect to a live catalog or database document what they need in their own .env.example — never real values.

Known gaps

Stated plainly rather than hidden behind a green badge:

• Tests exist only for telco-customer-intelligence so far (schema, referential integrity, no-leakage, causal signal, the churn case's leakage boundary and metric arithmetic, and the incrementality case's fence around its answer key — standard-library only, so CI runs them without installing anything). The migrated agent tracks still have none; CI's pytest job collects every tracks/<track>/<project>/tests/ directory as it appears. The test_*.py files under some scripts/ folders are manual integration scripts — they need a live catalog, a database and credentials — so CI deliberately does not collect them.
• data-privacy/ is empty; customer-analytics/ has its data model plus the churn and incrementality cases built, with three cases to go (see #64).
• Dependencies in the migrated agent projects are declared with ranges, not pinned. A fresh install today may not resolve to what the author ran. (The customer-analytics data model has no third-party dependencies — standard library only.)

License

MIT — see LICENSE.