I've been the person building KPI frameworks from scratch at three different startups. Every time it's the same thing: you set up tracking, write SQL queries, build dashboards, and then spend hours each week pulling numbers that should just be there. PMs at early-stage companies don't have a data team. They're doing this themselves, and most of the time the insights come too late to act on.
0 → 1 Product · 2025
Pulse
Sole Product Manager & Builder
An AI-powered analytics copilot that helps product teams at early-stage startups skip the SQL and get straight to the signals that actually matter.
PythonSQLLLM APIsNLPReactUser ResearchKPI Design
I interviewed 12 PMs and analysts at Series A through C startups. The pattern was consistent: they all had dashboards, but nobody trusted them to surface what mattered. 9 out of 12 said they still write ad-hoc SQL queries weekly because their dashboards don't answer the questions they actually have. The gap isn't data access. It's that the data doesn't talk back.
Tools like Amplitude and Mixpanel are powerful but designed for companies with dedicated analytics teams. Metabase and Mode are flexible but still require SQL fluency. The newer AI analytics tools (like Narrator, Zing) focus on visualization, not on the interpretive layer. None of them answer the question: "what changed this week, and should I care?"
I scoped the MVP around three high-confidence features using RICE: natural language querying (ask your data a question, get an answer without SQL), anomaly detection (flag when a metric moves outside its normal range), and weekly digest (an AI-generated summary of what changed and why it might matter). Pushed dashboards and custom alerts to v2.
Pulse connects to a startup's data warehouse (Postgres, BigQuery, or Snowflake). An LLM layer translates natural language questions into SQL, runs them, and returns plain-English answers with supporting charts. Anomaly detection runs nightly, flagging any tracked metric that moves more than 2 standard deviations from its trailing 30-day distribution.
The weekly digest combines that anomaly output with week-over-week deltas and a summarization model that writes like a PM's standup notes. All queries are read-only, and schema mapping happens once during onboarding so nothing needs configuring after setup.
- Time to insight: under 30 seconds. Median time from question submitted to answer viewed, against a baseline of about 25 minutes doing it manually.
- Weekly active query volume: up 40%. Lifted by defining activation, engagement, and retention metrics and running 3 A/B tests on the ML features behind the query experience.
- Activation rate (target: 60% in 24h): signups who connect a data source and run a first query within 24 hours.
- Query trust (target: 70%): queries where the user accepts the AI-generated answer without re-running it or switching to raw SQL.
- 4-week retention (target: 50%): activated users still running at least one query per week at the 4-week mark.
Why I built this
This wasn't hypothetical. I'd set up KPI tracking at Digo, built analytics frameworks at Giri, and automated data pipelines at UC Davis for over two years. Every time, I wished something like Pulse existed so I could skip the plumbing and get to the part that actually matters: making product decisions with real data.
Pulse shipped in 2025 and is live today. I'm still adding features based on usage data, and the targets above are what I'm tracking next.
← Back to selected work