Case 01 / 09

Scripts reviewed in days,not two weeks.

Two pipelines learn from real audience behavior and evaluate new scripts. AI advises, editors decide.

-80%Review cycle
100K→1MUsers during tenure
Head of Data & AIIdilio TV11/2025 to 02/2026
01 / Problem

Two weeks of writer round tables for every script.

The queue was the product bottleneck. Expert judgment lived in a room, while produced scripts, subtitles, and real audience behavior did not feed the next decision.

02 / Strategy

Instrument first, then build two pipelines over one knowledge base.

Pipeline A learns

It aligns produced scripts and subtitles with real audience behavior, joins performance, and mines evidence.

Pipeline B evaluates

It compares new scripts against that knowledge and gives editors a grounded report. AI advises, editors decide.

03 / Decisions

Cost, reversibility, and human judgment were architecture requirements.

  • Gemini and ChatGPT were compared and routed by cost, latency, and quality.
  • Prompts, embeddings, and features were versioned from day one.
  • Retries and alerts made failure visible.
  • Human review stayed in the loop by design.
04 / Results and under the hood

The room still decides. It stopped waiting.

Review fell from two weeks to days, an 80% reduction. Instrumentation and analytics guided acquisition with growth while the product moved from about 100K to about 1M users during the engagement.

Which opened the next problem

A million users write a million rows. The same patterns became a starter other teams can run.

Case 02 / 01

A production chatbot starter, in the open.

open sourceContinue the record
WhatsApp