4.2×
Increase in daily active users
ScaleTeam · Case Study
Natural Language Analytics Dashboard
ClaudePostgreSQLPythonReactPinecone
4.2×Daily active users85%Self-serve queriesMinutesvs. days for new dashboards
Chapter 01The Brief
The problem we
walked into.
Operators were waiting weeks for analysts to write SQL just to answer one-line questions.
A growing analytics SaaS had built a beautiful BI layer on top of a Postgres warehouse, but every non-trivial question still bottlenecked on the data team.
Customer success could not self-serve. Sales asked the same five questions every week. Product needed an answer faster than a sprint.
The team had tried off-the-shelf text-to-SQL — it hallucinated joins, ignored business definitions, and silently returned wrong answers.
Chapter 02The Build
How it
actually works.
Schema-grounded prompts, eval harness on real questions, and a confidence layer the UI actually surfaces.
- 01
Indexed schema + business glossary with embeddings (Pinecone).
- 02
Constrained generation with retrieved tables, columns, and metric definitions.
- 03
Built a 200-case eval harness driven by the actual analyst backlog.
- 04
Added a confidence model + explanation layer in the UI; low-confidence answers route to a human.
We treated the LLM as one component in a deterministic pipeline — not the pipeline. Retrieval handles correctness; the model handles synthesis.
Chapter 03The Outcome
Numbers that
moved.
Results we measured after 90 days in production — the metrics the client actually watches.
85%
Of queries now self-serve
Minutes vs. days
Time-to-answer for new questions
$2.1M
Net-new pipeline in first 90 days
“It didn't just save analyst time. It changed who in the company could ask the database a question — and that changed how fast we move.”
— — Head of Product, B2B Analytics SaaS
Ch 04Start Something
Your Product could be next.
Tell us about your project — We'll show you what's possible, what it costs, and what we'd ship first.
FREE DISCOVERY CALL. NO PITCH DECKS. REAL NUMBERS.