A detailed walkthrough of how a repetitive marketing audience request can move from manual analyst handling to an explainable, AI-assisted workflow—with query checks, sample output, auditability and human approval.
CRM teams frequently need an audience extracted from a data platform so a campaign can be launched. When every request requires manual analyst handling, turnaround and analyst capacity become bottlenecks.
A visual summary of how the request moves through the system—from a marketing need to a reviewed, auditable output.
Select any step to highlight its role. The sequence mirrors a practical implementation: request intake, historical pattern matching, query generation, validation and output update.
Capture request ID, campaign, audience description, inactivity window, country and exclusion rules from the request board.
Monday.com/APISchema validationCompare structured fields with historical requests and approved query patterns. Keep the match score and reason explainable.
Historical repositoryCursor / LLMFor a strong match, reuse approved logic. For a new pattern, draft a SELECT query and send it to review instead of blindly executing it.
SQL draftDatabricks GenieCheck query shape, exclusions, audience count and sample records. A designated reviewer verifies logic before production delivery.
GuardrailsHuman approvalSave the approved audience output, record an audit event, and comment/update the original request board.
RepositoryAudit logThe request is normalized into a consistent schema before any matching or query generation begins. Required fields and acceptable ranges are checked early so ambiguous requests do not silently progress.
{
"request_id": "CRM-DEMO-001",
"campaign": "Re-engagement",
"inactive_days": 30,
"country": "IN",
"exclude_flagged_users": true
}The most important decision is whether a request is sufficiently similar to a previously approved pattern. A partial match should not be treated as permission to auto-run a query.
Key criteria match an approved historical example. Reuse its query template, bind the current values, and still require output review.
Run static checks, preview audience counts, inspect exclusions, and prepare the request-board update only after checks pass.
The request differs on an important attribute, or the repository does not contain a sufficiently similar approved pattern.
Draft a query for analyst review. The reviewer confirms business definitions, exclusions, table/column logic and expected audience size.
This maps the business flow to the project structure in the accompanying Python repository.
Conceptual architecture, not a live integration map. In the demo scaffold, data and outputs are local; production connectors, warehouse execution, authentication and deployment remain separate implementation steps.
| Component | Responsibility | Production consideration |
|---|---|---|
app/main.py | API endpoints, request routing and dashboard serving. | Add authentication, authorization, rate limits and request correlation IDs. |
app/schemas.py | Validate required request fields and allowed value ranges. | Use explicit business rules and reject ambiguous or incomplete requests. |
app/matcher.py | Compare request attributes against approved historical patterns. | Calibrate thresholds on labelled examples; monitor false matches and drift. |
app/adapters.py | Demo query drafting, sample audience preview and request-board payload. | Replace placeholders with approved LLM, warehouse and Monday.com adapters. |
app/sql_guardrails.py | Basic static checks for the generated SQL draft. | Static checks are not a security boundary; enforce read-only access and table allowlists at the data platform. |
data/historical_requests.json | Stores example approved request patterns and templates. | Version templates, track owners/approvals, and avoid storing sensitive audience data. |
data/audit_log.jsonl | Records request status, match result and preview details. | Use durable, access-controlled logging with retention and privacy controls. |
The objective is to reduce repetitive work while preserving data quality, privacy and accountability.
Use operational metrics to demonstrate whether automation actually creates capacity and improves service levels.
Submit a sample request to the local API and see the matching result, SQL draft, audience preview and review steps.
For this interactive form to work, run the Python API from the project folder with uvicorn app.main:app --reload. The dashboard alone can still be read as a static explanation.
Submit the form to display the API response. This demonstration does not execute SQL against a live data platform.
-- The SQL draft will appear here