Music Distributor Data Assistant

Discovery & Requirements (Weeks 1-3)

We embedded with analytics, marketing, and A&R teams to map "business language" to the database schema. We cataloged 150+ frequently asked questions and conducted a comprehensive security and compliance review to understand royalty confidentiality and artist privacy requirements.

Agent Architecture Design (Weeks 4-6)

We prioritized specialization over generalization. No "god-mode" agent that tries to do everything. Instead, a hybrid "Two-Agent" system where each component does one thing well:

• The Brain (Context): Powered by OpenAI GPT-4o, managing conversation intent, memory, and formatting (e.g., drafting emails). Detects uncertainty and knows when to escalate.
• The Engine (Execution): Powered by Snowflake Cortex Analyst, keeping data processing inside the secure Snowflake boundary. Single-purpose SQL generation with clear handoffs.

Development & Integration (Weeks 7-14)

To handle the scale of an 80 TB warehouse, we built a production-grade stack:

• Frontend/Backend: A Next.js chat app with a Python/FastAPI orchestrator.
• Async Pipeline: Utilized Redis and ARQ workers to handle long-running analytical queries without blocking the UI.
• Initial POC: We first produced a Proof of Concept (POC) using Streamlit within Snowflake to validate the Cortex engine before moving to the custom app.

Testing & Refinement (Weeks 15-18)

We "red-teamed" the model against sample verified questions (e.g., Week-over-Week stream changes). We refined the Semantic Model (YAML) and enforced a strict "Single Query" rule in the system prompt to prevent inefficient, iterative database "chats" that bloat compute costs.

Performance Budgets & Guardrails

We set ceilings: max 1 tool call per query, 2-second latency budget, strict constraints on warehouse compute. The system exposes its reasoning path so we can catch tool-thrashing loops early. If the agent starts iterating inefficiently, we see it in the traces before it hits production costs.

Leveraging Existing dbt Intelligence

A critical accelerator was the client's dbt (data build tool) ecosystem. Their dbt view models already contained the business logic, joins, and column descriptions needed. By harvesting this metadata, we rapidly bootstrapped the Semantic Model (YAML), ensuring the AI utilized the same "source of truth" as the company's official dashboards.

Prioritizing the "Golden" Data

Not all tables are equal. We used endorsed statuses to signal which tables the agent should prioritize and built semantic models validated against existing dashboard outputs for high-stakes metrics like royalty payouts and stream counts. The agent knows which sources are authoritative and which are draft or deprecated.

Human-Readable Column Hygiene

We applied a simple test: if a junior analyst can't understand a column from its description, the AI won't either. Every table and column in the semantic model got plain-language descriptions. "net_rev_usd" became "Net revenue in US dollars after platform fees and distributor share, calculated at the end of each royalty period."

Architecting the "Brain" (Context Agent)

The GPT-4o "Brain" acts as an intelligent filter. Its system prompt handles:

• Intent Mapping: Distinguishing between a data query and an action request (e.g., "Draft an email summary").
• Constraint Enforcement: Explicitly forbids multi-step reasoning to maintain sub-2-second latency.
• Persona: Defined as a "Lead Music Industry Analyst" to ensure the use of standard terminology.

Snowflake Cortex & Intelligent SQL Generation

By utilizing Snowflake Cortex Analyst, the LLM generates SQL inside the Snowflake boundary. This ensures sensitive data never leaves the client's perimeter. The system achieves median translation latency under 500ms through pre-compiled templates and schema-aware prompt engineering.

Dynamic Visualization & Action

The system doesn't just return rows; it detects data shape to return trend lines or bar charts and can instantly "Draft an executive email" based on the results.

Self-Serve Data Access

Teams across Marketing, A&R, Finance, and Operations now query data independently for common questions. Requests that previously required analyst tickets and multi-day turnarounds are answered in seconds. Complex or novel queries still escalate to the analytics team, but routine data gathering no longer requires technical intermediaries.

Sub-2-Second Insight Latency

The system delivers median latency of 1.5 seconds from natural language input to visualization. Typical breakdown: 400-500ms for NL-to-SQL translation and 1-1.5s for SQL execution via optimized warehouses, depending on query complexity.

Zero SQL Required

The natural language interface understands complex queries like "Compare streaming growth in Latin America vs. Europe for artists signed in 2023" and translates them into optimized multi-table joins.

Team Impact & Efficiency

Comparing ad-hoc ticket volume before and after deployment, the analytics team saw a 70% reduction in routine query requests, freeing capacity for strategic predictive modeling. In internal surveys, business teams reported significant time savings on day-to-day data gathering tasks.

Operational Efficiency

The system eliminated the need for additional BI tool seat licenses and reduced analyst hours spent on routine queries, contributing to measurable operational cost savings.