General-purpose LLMs scored below 52% on multilingual ecommerce tasks. Shopee's own 245B-parameter Compass now handles 340B tokens monthly.
Southeast Asia's leading ecommerce platform, serving hundreds of millions of monthly active users across eight markets and facilitating over $136 billion in annual gross merchandise value.
General-purpose LLMs performed well on English tasks but degraded sharply on the specialized, multilingual work that defines ecommerce across the...
“Going from 3 billion to 340 billion tokens in just eight months wasn't merely a scaling story—it validated our ability to operationalize AI at platform scale. By leveraging a diverse portfolio of models running on NVIDIA GPU infrastructure, including our in-house Compass model, we've been able to rapidly expand AI adoption across the business while delivering the performance and reliability required to support our growing platform.”
E-commerce marketplace and mobile shopping platform.
NVIDIA is a technology company that specializes in semiconductors, graphics processing units, and artificial intelligence for applications in data centers, gaming, and more.
Shopee's Multilingual ecommerce model is part of this use case:
Related implementations across industries and use cases
Complex forms stalled sellers. Multimodal AI now drafts listings from video, cutting setup time to minutes.
Manual descriptions and secondhand slang bottlenecked new sellers. Now, AI generates listings and translates niche recommerce terms.
Keyword search failed on complex vehicle queries. Now, an AI pipeline rewrites messy questions to power shopping and service assistants.
Rate limits capped geocoding at 2,000 requests/min. A fine-tuned LLM now clears 8,000/min—cutting costs by 80%.
Processing 2T tokens on CPUs took days. GPU acceleration cut prep to hours, unlocking a 5% accuracy gain.
Workflows collapsed when staff changed roles. A meta-agent now turns plain-language descriptions into live agents—no AI team needed.
An aging homegrown CRM held back 1,700 agents. Now, an AI co-pilot helps them draft faster, higher-quality emails from a single screen.
Protecting users from harmful on-device AI required internet. A powerful safety AI now runs directly on the PC, guarding users even when offline.
A 200% yearly data expansion bottlenecked global operations. Now, AI accelerates coding, drafts recipe cards, and resolves inquiries.
General-purpose LLMs scored below 52% on multilingual ecommerce tasks. Shopee's own 245B-parameter Compass now handles 340B tokens monthly.
Southeast Asia's leading ecommerce platform, serving hundreds of millions of monthly active users across eight markets and facilitating over $136 billion in annual gross merchandise value.
General-purpose LLMs performed well on English tasks but degraded sharply on the specialized, multilingual work that defines ecommerce across the...
“Going from 3 billion to 340 billion tokens in just eight months wasn't merely a scaling story—it validated our ability to operationalize AI at platform scale. By leveraging a diverse portfolio of models running on NVIDIA GPU infrastructure, including our in-house Compass model, we've been able to rapidly expand AI adoption across the business while delivering the performance and reliability required to support our growing platform.”
E-commerce marketplace and mobile shopping platform.
NVIDIA is a technology company that specializes in semiconductors, graphics processing units, and artificial intelligence for applications in data centers, gaming, and more.
Shopee's Multilingual ecommerce model is part of this use case:
Related implementations across industries and use cases
Complex forms stalled sellers. Multimodal AI now drafts listings from video, cutting setup time to minutes.
Manual descriptions and secondhand slang bottlenecked new sellers. Now, AI generates listings and translates niche recommerce terms.
Keyword search failed on complex vehicle queries. Now, an AI pipeline rewrites messy questions to power shopping and service assistants.
Rate limits capped geocoding at 2,000 requests/min. A fine-tuned LLM now clears 8,000/min—cutting costs by 80%.
Processing 2T tokens on CPUs took days. GPU acceleration cut prep to hours, unlocking a 5% accuracy gain.
Workflows collapsed when staff changed roles. A meta-agent now turns plain-language descriptions into live agents—no AI team needed.
An aging homegrown CRM held back 1,700 agents. Now, an AI co-pilot helps them draft faster, higher-quality emails from a single screen.
Protecting users from harmful on-device AI required internet. A powerful safety AI now runs directly on the PC, guarding users even when offline.
A 200% yearly data expansion bottlenecked global operations. Now, AI accelerates coding, drafts recipe cards, and resolves inquiries.