We use cookies
To improve your experience. Cookie policy
DataPulse · Business Intelligence · 2024
DataPulse wanted to let non-technical customers query their data warehouse in plain English. We built a RAG-powered analytics layer that translates natural language questions into accurate SQL — with explainability built in.
Key outcomes — Business Intelligence
71%
30-day retention
from 34%
47s
Time-to-insight
from 18 minutes
96%
Query success rate
at 14K+ queries/day
$42M
Series B valuation
closed post-launch
§ 01 / Problem
DataPulse's customers were churning after 30 days because their analysts couldn't get value from the product without SQL skills. The existing query builder required 12 interactions for a question a data analyst would answer in 30 seconds. Product leadership had 6 weeks before their Series B investors expected a demo of the natural language feature.
§ 02 / Approach
The key insight from our initial discovery was that schema diversity — not model capability — was the binding constraint on accuracy. A query that worked on one customer's warehouse would fail on another's because column names and table structures varied wildly. We built a schema-aware embedding index that captures table names, column descriptions, and sample values, then retrieves the most relevant context before any LLM query. The first two weeks produced a benchmarkable prototype; the following four weeks were iterative accuracy improvement through prompt engineering, schema annotation tooling, and a curated test suite of 500 representative queries.
§ 03 / Solution
A Retrieval-Augmented Generation pipeline using GPT-4o for query intent classification, a schema-aware embedding index for table and column disambiguation, and a validated SQL generation layer that runs generated queries against a sandboxed read replica before returning results. The React frontend shows both the generated SQL and an explanation of the query logic — building user trust and enabling learning.
§ 04 / Result
30-day retention increased from 34% to 71%. Average time-to-first-insight dropped from 18 minutes to 47 seconds. DataPulse closed their Series B at a $42M valuation, citing the NL analytics feature as the primary differentiator. The feature now handles 14,000+ queries per day with a 96% query success rate.
71%
30-day retention
from 34%
47s
Time-to-insight
from 18 minutes
96%
Query success rate
at 14K+ queries/day
$42M
Series B valuation
closed post-launch
Tech stack
The natural language analytics feature was the centrepiece of our Series B pitch. stackloader delivered it in six weeks, it worked correctly 96% of the time out of the gate, and it looked beautiful. We closed $42M partly because of what they built.