The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
Title: Customer-Facing AI Analytics Chat on Snowflake
# Problem
20K+ business customers wanted answers about their own business without building dashboards or waiting on support. Pointing an LLM straight at raw tables gave wrong joins, inconsistent metric definitions and numbers that didn’t match the product’s dashboards, so nobody trusted the answers.
# What I did
- Explored options for Semantic layers and integrated dbt Smentic layer to company stack.
- Defined business metrics once in the dbt Semantic Layer, the single source of truth for internal, external analytics (this grew eventually to ~300 metrics)
- Built a chat interface where the LLM picks governed metrics and dimensions from the semantic layer instead of writing free-form SQL.
- Kept it multi-tenant and secure: each customer only ever sees their own data
Built an evaluation framework to measure and protect accuracy:
- A golden set of real customer questions, each with the expected metric, dimensions, filters and answer
- Covers frequent lookups, time comparisons, filtered breakdowns and multi-metric questions
- Automated scoring: correct metric and filter selection, plus result match against the expected answer.
- Separate LLM agents evaluates the answer in predefined criteria and steps.
# Result
- 92% of eval questions answered correctly
- Answers in under 10 seconds
- 300+ governed metrics available in plain English to 20K+ customers
- Same numbers in chat as in the dashboards; new metrics show up in chat as soon as they’re added to dbt.
# Tools:
- Snowflake, dbt Semantic Layer, Kimball dimensional modeling, LLMs, Python
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Related posts
Game Player Data Pipeline and Dashboard
AI-Powered Facebook Messenger Follow-Up Automation | n8n
I built an AI-powered Facebook Messenger Follow-up Automation workflow using n8n to help businesses reconnect with potential customers and reduce missed sales opportunities.
When a customer asks about a product but stops replying, businesses often lose track of the conversation. This workflow helps automate timely, personalized follow-ups while keeping customer experience in mind.
What the Workflow Does
Automated Follow-ups: Sends follow-up messages at the right time.
AI-Powered Personalization: Uses AI to generate personalized messages based on customer conversations.
Order Verification: Checks order status before sending reminders.
Communication Scheduling: Avoids sending messages during late-night hours.
Sales Support: Helps businesses reconnect with interested buyers and reduce manual follow-up tasks.
Tools & Technologies
n8n | AI | Facebook Messenger | Workflow Automation | Order Verification
Project Goal
Demonstrate how AI-powered automation can improve customer communication, streamline sales follow-ups, and create more opportunities for businesses to convert interested buyers.
This is a demo workflow showcasing the potential of AI automation for eCommerce businesses.
Services
I help businesses build AI-powered workflows, sales automation systems, customer support automations, and custom n8n integrations.
#n8n #AIAutomation #FacebookMessenger #eCommerce #SalesAutomation #WorkflowAutomation #Contra
The late-night quiet hours are a nice touch. When a follow-up gets held for the morning, does the workflow re-check the order status at send time, or only when the reminder is first queued?
An AI agent answers questions from real store data. A check layer blocks numbers that AI makes up.
For that check layer, I'd include a test where the answer uses a real store number with the wrong metric or date range. How do you tie each numerical claim back to the source result and its filters?