Data questions to BI fell about 90% after I rebuilt it. I help other teams free up time the same way.

Both systems below run in production at a 90-person HR tech company, where I lead BI in-house. For other companies, I advise, build it with the team, or teach them to build it themselves.

Book a 30-minute callSee the Starter Kit

90-person HR tech company · in-house BI rebuild · 9 months live

I rebuilt the BI function so the company answers its own data questions.

My role: I designed and built the system, and wrote the rules it runs on.

  • 90%fewer data questions reaching BI: from about 20 a week to 1 or 2
  • 10+people outside BI who now get their own answers
  • 0new hires needed, down from the 1 the team had originally planned
Before
Scattered docs, disconnected tools, and knowledge in people's heads meant every question waited in the BI queue.
What I built
A system where agents answer stakeholders' data questions, catch breaking changes before they ship, and keep the company's data knowledge current every week.
What keeps it reliable
Written rules: a constitution, a fixed structure, and defined workflows. My teammate works through the same system every day.

How it works

Each question, check, and update in the system runs through one shared knowledge base.

Data sources, Amplitude, Refiner, HubSpot, Stripe, Segment and BigQuery, connect to Hex. The knowledge base in git, 405 verified entries, is compiled into the Hex agent context of 106 guides, and stakeholders ask Hex agents in plain language. Each week, gaps found in agent chats are checked, approved, and added back to the knowledge base. 33 agent workflows, 11 on a schedule, run on the knowledge base and catch breaking changes and tracking drift.

90-person HR tech company · marketing attribution build

I built the system that shows which channels drive signups, and what they cost.

My role: I designed and built the stack; my teammate wrote parts of the attribution SQL.

  • 0.1% → 95%signups with a known channel
  • 47expert sources on attribution I gave the AI agents to learn from before building
  • 100%of acquisition cost allocated to channels: ad spend plus people, tooling, and overhead
Before
Nearly every signup landed in “Direct”, so nobody could see what a channel cost or what a campaign returned.
What I built
A channel classifier and 13 daily SQL models that attribute every signup, then price each channel with ad spend and finance costs.
What changed
Finance, leadership, and marketing now read channel cost and campaign return from one place, and Google Ads learns from attributed signups.

How it works

Each signal from ads, forms, and product data flows into one system that prices each channel.

The website sends consent and tags through Google Tag Manager to Amplitude events and the signup survey. An attribution model in BigQuery, a channel classifier and 13 daily SQL models, raised signups with a known channel from 0.1% in January 2026 to 95%, part tracked directly and the rest estimated from the survey. Cost allocation adds ad spend and finance costs, and Hex dashboards show CAC by channel and campaign return. Attributed signups flow back into Google Ads as the conversion signal. Before building, the AI agents learned from 47 expert sources on attribution.

Also applied to

  • Finance

    The workspace I built survived a handover; the incoming owner ran the recurring work alone after 2 sessions, and it's still in use months later.

  • Product marketing

    Emails used to need manual fixes between the design tool and delivery platform; the next one rendered correctly on the first attempt, in under 1 hour.

Let's build something like this for your team.

Working together can mean advisory, hands-on implementation, or training your team to build it themselves. If you want to start smaller, the €300 Starter Kit gives you a working workspace, 6 installable skills, and a 2-hour session.

Book a 30-minute callSee the Starter Kit