Skip to content
Finalist

Data Contract Governance Agent

Data Contract Governance Agent is an AI agent on Snowflake that onboards new data products safely.

Code Video Cerebral Valley

Core trick

Routes contracts through deterministic rule checks, has the LLM draft only template-based fixes, and applies Snowflake masking policies only after a named steward approves.

Tags

dev toolshealth

Description

Data Contract Governance Agent is an AI agent on Snowflake that onboards new data products safely. A data owner submits a data contract. The agent routes it to the right governance track (standard, financial or regulated lab), lists every gap with its rule ID, checks whether the data was actually written by an instrument, a person or an AI, and drafts the fixes. A named data steward approves, and only then are masking policies and tags applied and the data promoted to production. The problem: onboarding a regulated data product takes about three weeks (our estimate) of back-and-forth across lab, governance, security and engineering teams. Owners and SLAs are missing, patient IDs are unmasked, and a new risk has appeared: AI-generated values can enter pipelines labeled as instrument data, with nobody checking. Design principle: the AI drafts, fixed code enforces, and people decide. Routing is deterministic, fixes come only from vetted templates (Claude never writes SQL), and the agent can never approve its own work. Every action is logged with the model, agent and rule version. All data is synthetic. The agent governs how data is handled and never interprets lab values.
ashwajeetbetal/data-product-onboarding-agent★ 00 forksPythonCreated Oct 23, 2026 · last push Oct 23, 2026

README excerpt

# Data Contract Governance & Onboarding Agent

A Cortex Agent that reviews the data contract for a new data product, routes it to a governance track
(STANDARD, FINANCIAL or REGULATED_LAB), lists its gaps with rule IDs, drafts fixes from vetted templates, and applies
them only after a named data steward approves. Built on Snowflake for the Snowflake hackathon (Fri Oct 23).

**For whom:** the **lab data steward** at a fictional clinical lab. Today the steward reads each new contract by hand,
checks it against the rulebook, chases the owner for missing fields, and finds out late that a batch was written by an
AI imputation job. This agent does the review in minutes, leaves every decision to the steward, and records it.

Synthetic data only: no PII, no real company, no proprietary knowledge. The agent governs how data is handled; it never
interprets lab values or gives health guidance.

**Video:** _<link to the one-minute demo video>_

## What it does (the C-LAB-001 demo)

1. The lab data owner (USER2) asks the agent in **Snowflake CoWork** or in **Claude via MCP** to review contract C-LAB-001.
2. The agent runs the review in a fixed order: parse the YAML with code, profile the columns (no raw values),
   route the track deterministically, check provenance against WRITE_LOG, search the rulebook, propose fixes.
3. It reports **REGULATED_LAB** and four gaps: STD-01 (no owner), STD-02 (no SLA), LAB-01 (patient_id not protected)
   and LAB-04 (batch **LAB-B010**, 50 rows, written by

Team