Valentin Galfré

2025 Platform and AI

A content engine that writes 90% of Coderhouse

Staff Engineer, technical lead for platform engineering and AI

Coderhouse is one of Latin America's largest edtech platforms, with 500k active students. Since September 2025 I lead its platform engineering and AI initiatives, and the biggest piece of that is the system that now produces most of the platform's educational content.

Problem

Producing educational resources for hundreds of courses was slow and manual, and infrastructure had to hold up under peak demand of tens of thousands of concurrent students.

Solution

I designed and shipped multi-agent AI systems that generate resources across text, images and audio, with structured outputs and evals so quality is measured rather than assumed. In parallel I defined a distributed infrastructure architecture built for resilience and high availability.

Outcome

Thousands of educational resources are generated automatically every day, and the platform serves 50,000+ concurrent users through peak demand. Engineering standards around reliability, maintainability and performance were adopted organization-wide.

Impact

The AI systems power 90% of the platform's content production.

A classical painting of an open-air academy where a fountain releases hundreds of glowing pages into a stormy sky while students reach up for them.
A fountain that writes. Most of the platform's content comes out of it.
0 active students on the platform
0 concurrent users at peak
0% of published content produced by the system

How a resource gets made

A brief comes in: a course, a lesson, what the student needs to walk away with. A planner agent turns it into a structure. Writer, illustrator and narrator agents produce text, image and audio in parallel. Nothing is published until it passes evals that score it against the brief.

brief planner writer illustrator narrator evals published
One brief in, one resource out: text, image and audio produced in parallel, measured before anyone sees them. This runs thousands of times a day.

What it took to trust it

The hard part of a system that writes most of what a platform publishes is not generation. It is knowing when the output is good without a person reading every piece. Structured outputs made every result checkable. Evals made quality a number that trends instead of an opinion that varies.

The infrastructure underneath had to hold the same standard: designed for peak demand, resilient by default, observable enough that a regression is a ticket for one team and not a meeting for five.

Around the work

Coderhouse ↗