Case study · AI product

account health, live.

How I turned my team's slowest morning ritual — pulling numbers by hand — into a dashboard that already knows.

Shipped Feb 2026 Bidease · internal tool Users: CSM + AdOps Claude Code · MCP · LLM

the problem.

Every account question started the same way: someone opens the DSP, exports numbers, pastes them somewhere, and repeats it tomorrow. Across ~8 UA clients in two markets, those recurring manual pulls ate the hours meant for actual strategy — and the data was stale by the time anyone read it.

what I built.

An LLM-powered account-health dashboard connected to live internal DSP data through Model Context Protocol (MCP), refreshing every five minutes. It centralizes prioritized day-over-day, week-over-week, month-over-month, and quarter-over-quarter analysis, so the team reads answers instead of assembling them.

Recreated and anonymized for this page — illustrative numbers, real product shape.

what changed.

Zerorecurring manual pulls — the dashboard already has it.
5 minutesfrom DSP reality to the numbers the team reads.
Capacity backhours returned to strategy instead of spreadsheet assembly.

It also set the stage for what came next: automated Slack alerts on pacing, KPI variance, budget scaling, creative performance, and fraud signals — the dashboard that answers became the system that speaks up first.

how I work.

  1. DiscoverySit with CSM and AdOps, map what they pull by hand and when it hurts.
  2. DesignInterfaces, workflows, and prompt design — what the team sees and asks.
  3. BuildAI-assisted development with Claude Code; MCP wires in live DSP data.
  4. ValidateManual source-data checks; test outputs for hallucinations and logic errors.
  5. AdoptTeam training until it is the default tab, not the demo.
Back to impact The résumé