Speakers: Tom Isles (Canva)
Track: Software Engineering
Canva's "Ingredient Generation" service powers all of our media generation experiences. Every image, video, audio and 3d object request for generation goes through this service, which offers 50+ AI Models. One of its core capabilities is the ability to automatically switch between models when one fails.
As the service has grown, product logic specific to individual models got tangled up with inference logic. GPT Image 1 could tailor its image to a users design aesthetic, while Gemini couldn't. When the system failed over to a back up model, features the backup didn't support would silently break - no error, just a bad result.
Fixing it meant migrating 70,000 lines of code, to extract the feature logic from the inference logic into separate packages. But, the complexity was too high to do it in a single pull request - it needed to be done gradually, with care.
The migration touched every live AI request on the platform. One wrong step and it's a major outage. Pre-AI, this might have taken months. A team of 3 did it in just over a month, using AI.
This talk is about migrating 70,000 lines of live code without an outage using AI. It covers why the riskiest decisions stayed manual, how agent mistakes were turned into reusable skills, and how AI was used to ensure a safe and smooth rollout that resulted in a successful migration - without a single bug introduced in the process.
https://data.webdirections.org/ai-engineer-sydney/sessions.json#25ab91c6-90a9-44dc-9b3a-bea252dfc420
Full text on webdirections.org
.