Uploaded May 2025 | Updated September 2026, 2 weeks ago
Learn how Mirakl transformed one of e-commerce's biggest pain points—catalog onboarding—from a weeks-long process to an automated, AI-powered system using fine-tuned large language models. This technical case study explores the Catalog Transformer, a production system that revolutionizes how marketplace sellers can quickly get their products online, directly impacting business growth and seller success.
This talk covers the architecture and operational choices behind Mirakl’s production LLM platform. Learn how they integrated LoRA adapters for efficient updates, deployed elastic autoscaling for traffic bursts, and served models dynamically on Anyscale using LLaMa 3.1 8B. Their solution processes millions of predictions monthly while keeping costs under control.
This talk is essential for ML engineers, platform architects, and product teams working on production LLM applications. Get practical insights into architecture choices, lessons learned from productionizing language models at scale, and cost optimization strategies that enable sustainable AI-powered business processes in the competitive e-commerce landscape.
00:00 - 01:15 - Who is Mirakl
01:16 - 04:57 - Challenges with e-commerce catalog onboarding
04:58 - 09:46 - Scaling with LLMs, cost and other considerations
09:47 - 15:34 - PoC to production requirements
15:35 - 19:47 - PoC results
19:48 - 23:48 - Production metrics and architecture
23:49 - 25:03 - Next ... multimodality
--
Liked this video?
Check out the other presentations: youtube.com/playlist?list=PLzTswPQNepXlWVHzNpyWxyvyzeOV-5-uA
--
🔗 Connect with us:
- Subscribe to our YouTube channel: youtube.com/@anyscale
- Twitter: https://x.com/anyscalecompute
- LinkedIn: linkedin.com/company/joinanyscale
- Website: anyscale.com
Learn how Mirakl transformed one of e-commerce's biggest pain points—catalog onboarding—from a weeks-long process to an automated, AI-powered system using fine-tuned large language models. This technical case study explores the Catalog Transformer, a production system that revolutionizes how marketplace sellers can quickly get their products online, directly impacting business growth and seller success.
This talk covers the architecture and operational choices behind Mirakl’s production LLM platform. Learn how they integrated LoRA adapters for efficient updates, deployed elastic autoscaling for traffic bursts, and served models dynamically on Anyscale using LLaMa 3.1 8B. Their solution processes millions of predictions monthly while keeping costs under control.
This talk is essential for ML engineers, platform architects, and product teams working on production LLM applications. Get practical insights into architecture choices, lessons learned from productionizing language models at scale, and cost optimization strategies that enable sustainable AI-powered business processes in the competitive e-commerce landscape.
00:00 - 01:15 - Who is Mirakl
01:16 - 04:57 - Challenges with e-commerce catalog onboarding
04:58 - 09:46 - Scaling with LLMs, cost and other considerations
09:47 - 15:34 - PoC to production requirements
15:35 - 19:47 - PoC results
19:48 - 23:48 - Production metrics and architecture
23:49 - 25:03 - Next ... multimodality
--
Liked this video?
Check out the other presentations: youtube.com/playlist?list=PLzTswPQNepXlWVHzNpyWxyvyzeOV-5-uA
--
🔗 Connect with us:
- Subscribe to our YouTube channel: youtube.com/@anyscale
- Twitter: https://x.com/anyscalecompute
- LinkedIn: linkedin.com/company/joinanyscale
- Website: anyscale.com
