vasundhara 1.1.1

Synthetic multi-party dataset generator for data clean room testing

Dhara

Synthetic multi-party dataset generator for data clean room testing.

Reap what you sow.

---

What is Dhara?

Data clean room tests fail not because the infrastructure is wrong — but because the test data is wrong.

Nulls, mismatched IDs, unrealistic distributions, and uncontrolled overlap rates make clean room query results meaningless during development. Dhara solves this at the source.

Given a use case configuration, Dhara generates N CSV files — one per data partner — where:

• Each file has a match ID column — SHA-256 hashed (hashed_email, hashed_phone, hashed_customer_id) or UUID v4 (device_id / IDFA / GAID)
• Cross-file ID overlap is mathematically controlled via a pairwise match rate matrix (inclusion-exclusion algorithm)
• Column values follow your defined distributions — weighted categorical, integer, decimal, boolean, date, datetime, email, free text
• Nullable percentages are respected per column
• Achieved match rates are verified post-generation and reported in metadata.json
• Participant schemas can be imported from a pasted CSV sample or JSON column list — types, distributions, and value weights inferred automatically
• All runs are reproducible via a seed

---

Quickstart

Web UI (Docker)

docker pull pn3uma7/dhara
docker run -p 8080:8080 pn3uma7/dhara
# Open http://localhost:8080

CLI

pip install vasundhara
dhara init                                          # interactive config builder → usecase.json
dhara generate --config usecase.json                # generate datasets → ./dhara_output/
dhara generate --config usecase.json --seed a3f9    # reproduce an exact run

---

How it works — 5 steps

Step	What you do
1 Use Case	Name your scenario, choose match ID type, set number of partners
2 Participants	Import each partner's schema from a CSV sample or JSON, or build manually
3 Match Rates	Set pairwise overlap % — drag sliders, see absolute pool sizes live with industry benchmarks
4 Review	Run the Audience Simulator — see how filters reduce the shared pool before generating
5 Generate	Download a ZIP: one CSV per partner + metadata.json with match rate verification

---

Example use case

Three-way clean room intersection: Brand × Publisher A × Publisher B

Match rates:  Brand ∩ Pub-A = 20%,  Brand ∩ Pub-B = 15%,  Pub-A ∩ Pub-B = 10%
All-3 overlap: 5%

Brand query:    WHERE segment = 'premium' AND recency = 'active'
Pub-A query:    WHERE content_vertical = 'finance'
Pub-B query:    WHERE device_type = 'mobile'

→ Shared pool before filters: ~150,000 IDs
→ After all three filters:    ~18,000 – 25,000 IDs

See examples/monster_doordash_walmart.json for a complete ready-to-run config.

---

Features

Match ID types

Type	Format	Use case
hashed_email	SHA-256 hex	Email-based clean rooms
hashed_phone	SHA-256 hex	SMS / mobile targeting
device_id	UUID v4 (IDFA / GAID)	Mobile app audiences
hashed_customer_id	SHA-256 hex	CRM / loyalty ID matching

Column types

Categorical · Integer · Decimal · Boolean · Date · Datetime · Email · Free Text

Schema inference

Paste a CSV sample → Dhara infers column types, value distributions, and weights from actual frequency data. Match key columns are auto-detected and skipped.

Audience Simulator

Before generating, simulate a clean room query: apply one filter per partner and see how many IDs survive — with a per-step funnel and survival rate.

Proportional weight sliders

Drag a categorical value's weight up — others reduce proportionally. Hard 95% cap prevents any one value from owning the entire distribution.

Reproducibility

Every run produces a seed. Re-run with --seed to get byte-identical output.

---

Output

dhara_run_20260530_143022.zip
├── brand_dataset.csv           ← match ID + column values
├── publisher_a_dataset.csv
├── publisher_b_dataset.csv
└── metadata.json               ← seed, requested vs achieved rates, drift warnings

---

Tech stack

Layer	Technology
Frontend	React 18, TypeScript, Tailwind CSS, Ubuntu font
Backend	Python 3.11, FastAPI
Data generation	NumPy, Pandas, Faker
ID generation	hashlib SHA-256, uuid v4
Packaging	Docker · PyPI (pip install vasundhara) · typer CLI

---

Self-hosting

# One-time setup — no auth, no data leaves your machine
docker run -d -p 8080:8080 --restart=always pn3uma7/dhara

# Expose via VPN or internal DNS: dhara.yourcompany.com

---

Part of the Vasu tool universe. Dhara (धरा) — earth, foundation, the one who holds.

MIT Licence — free to use, self-host, and modify.