The Fifth Elephant
The Fifth Elephant

The Fifth Elephant 2026 Winter Edition

Call for Data Problems and Proposals

All submissions

Satyavrat Bondre

Show me the rows: metrics in Claude, not text-2-sql chatbots

Submitted Oct 11, 2026

🎤 I am submitting: To speak I have a submission for: Showcase - a cool idea I have built/working on - 20 mins

Submission type

Talk/session proposal - 30-40 mins

If talk/session proposal, showcase, or poster
Describe your session in 2 paragraphs.

Our finance team needed answers from the data warehouse without waiting on the data team. How much ARR do we have? How many deals closed this month? What did renewals look like on this date last year? The usual approach to “AI over the warehouse” is to hand a model the data dictionary and the raw tables and let it write SQL. We tried that first, and it gave confident numbers that matched nothing finance trusted. The warehouse holds raw source-system tables that look perfectly usable and reconcile to nothing. Two columns are misspelled in the warehouse itself. One dimension table has duplicate rows per account, and IDs come in 18-character and 15-character versions that don’t join. Even our own hand-written queries disagreed with the dashboards: off by 1,755 contracts on ARR, and showing 3 new-business deals for the month where the dashboard showed 228. More documentation wasn’t going to fix that.

So we made the model’s world smaller, and we didn’t build a chatbot to do it. The model gets a catalog of about two dozen certified metrics instead of tables and a dictionary. Each metric is one query that matches one number on a dashboard people already use, and it works for any past date too. Every answer comes back as an Excel file with the exact rows and SQL behind the number. If a question falls outside the catalog, the assistant says so and lists what it can answer. The finance team already used Claude every day, so we skipped the custom chat app and the months of login, UI and hosting work it would have taken. None of that would have made a number more correct. The whole product is a folder of instructions plus one script that runs on the user’s laptop over the VPN, and all that saved effort went into the metrics. We’ll walk through the catalog, how we certified each metric, and what we’re still unsure about.

Mention 1-2 takeaways from your session.

  1. For questions about numbers, use a small set of certified metrics that each reconcile to something people already trust, and ship the rows with every answer. That works better than a full data dictionary plus raw tables.
  2. Before building a chatbot for an internal team, check whether the assistant they already use can do the job with good instructions and one safe tool. If it can, put the effort into the data instead of the interface.

Which audiences will benefit most from your session?

Analytics and data engineers who are being asked to “put AI on the warehouse”. Finance and RevOps people who want self-serve numbers they can actually check. Anyone deciding between building an AI app and configuring an assistant their team already has.

Add your bio: who you are, where you work, and any relevant context.

I’m Satyavrat, and I work as a consultant (often with Dono Consulting as their Head of Engineering).

Optional: What don’t you know yet, or what do you want help with?

  • Growing the catalog without losing trust. Every new kind of question needs a new certified query. How do others decide what gets certified, and who owns a metric when the dashboard behind it changes?
  • Where the line sits. Some questions are one small step past a certified metric, like a different grouping or a filter. We currently mark those as uncertified and say so. We’d welcome critique on whether that’s too strict or not strict enough.
All submissions

Comments

{{ gettext('Login to leave a comment') }}

{{ gettext('Post a comment…') }}
{{ gettext('New comment') }}
{{ formTitle }}

{{ errorMsg }}

{{ gettext('No comments posted yet') }}

Hosted by

The Fifth Elephant
The Fifth Elephant
Jumpstart better data engineering and AI futures