Problem, approach, result. With a number.

Problem

Coastal Community Bank had no analytics engineering practice. Reporting logic lived in ad hoc SQL and analyst tribal knowledge. Standing up a dbt foundation the traditional way was scoped at a full quarter, and the team could not pause its reporting load that long.

Approach

I founded the bank’s first dbt project: database architecture, development standards, and best-practice guidelines written before the first model. Then I put Codex coding agents inside that guardrail. Agents drafted staging and intermediate models from documented source contracts; I ran every output through the same QA and code review bar a human analyst would face. The agents did the typing; the standards did the governing.

Result

The initial data model build, scoped at one quarter, shipped in one week: roughly 13x faster. The standards and training I authored mean the analyst team now extends the project without me in the loop.

Problem

The analytics team spent roughly 90% of its time on ad hoc requests. That is a team functioning as a human query interface, with no capacity for the strategic work it was hired for.

Approach

Rather than another dashboard mandate, I introduced AI analytics agents as the self-service layer. The playbook:

1. 1 Build a governed dbt semantic layer so the agent queries certified definitions, not raw tables.
2. 2 Pilot with the highest-volume request categories.
3. 3 Publish metric dictionaries so agent answers are auditable.
4. 4 Route what agents cannot answer to analysts as properly scoped projects, not interruptions.

Result

The strategy is projected to cut non-value-add ad hoc workload from about 90% of team time to as low as 15%, converting most of the team’s calendar from reactive queries to leveraged analytical work.

Problem

At Dolly, Retail Ops reschedules were eroding partner trust and revenue. Intuition-driven fixes were not moving the number.

Approach

I built a decision-support tool that put the right data in front of ops at decision time, then led the experimental design to measure impact honestly: clean control groups, pre-registered metrics, and enough runtime to trust the readout. The same discipline ran the company A/B testing program, including the test series that cut apartment-move cancellations by 15%.

Result

Reschedule rate down 50%. Partnership volume up more than 50%. Measured, not estimated.