Customer Service Automation · Fraud Detection · Marketing Personalization
Shopee needed large language models that could handle multilingual ecommerce tasks across Southeast Asia, Taiwan, and Brazil, but general-purpose frontier models degraded sharply on regional languages, shopping concepts, fraud workflows, and customer-service use cases.
Reported impact
Shopee scaled Compass API usage from 3 billion to 340 billion monthly tokens in eight months, delivered 50x faster fraud detection at 90 percent lower cost, outperformed frontier models on regional ecommerce benchmarks, and expanded AI across conversion, ads, logistics, and multilingual customer interactions.
Shopee built its own ecommerce LLM stack, centered on the Compass family of models, on top of NVIDIA AI Factory infrastructure. The company pretrained Compass-v3 as a 245B-parameter mixture-of-experts model on 12 trillion tokens of curated multilingual corpora and synthetic ecommerce instructions, using NVIDIA Megatron-Core for distributed pretraining across NVIDIA A100, H100, and RTX PRO 6000 Blackwell GPU fleets. In post-training, Shopee uses NVIDIA NeMo for instruction tuning, preference alignment, and agentic reinforcement learning methods such as OTPO and SPEC-RL. NVIDIA TensorRT-LLM then serves Compass models in production across search, recommendations, fraud detection, customer service, and logistics. This gives Shopee one domain-built platform that connects large-scale training, multilingual alignment, and production inference for regional commerce workloads.
Shopee curates multilingual ecommerce corpora and proprietary marketplace data into the Compass training pipeline, Megatron-Core and NeMo produce aligned domain models on large GPU fleets, TensorRT-LLM serves those models in production, and the outputs feed search, recommendations, fraud, logistics, and customer-service experiences across regional markets.
At inference, NVIDIA TensorRT-LLM serves the Compass models in production.
For the post-training phase ... the team uses the NVIDIA NeMo framework for instruction tuning and preference alignment.
Monthly API tokens grew from 3 billion to 340 billion in eight months—a 113x increase.
NVIDIA Megatron-Core handles distributed pretraining at scale.
Shopee’s AI team pretrained Compass-v3, a vertical-domain mixture-of-experts model with 245B total parameters and 71B active per token.
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
Google Cloud Sample Modified · Oct 9, 2026 Central Retail Food Replaced ~2,000 monthly manual Excel reports with a single source of truth on BigQuery through Looker dashboards Central Retail Food unifies its ecosystem on Google Cloud, moving from fragmented data to conversational… Addresses 50% to 60% of customer service queries with automated chatbots built on Conversational Agents BigQuery · Gemini · Gemini platform 
In-house Sample Sep 3, 2026 Grupo Falabella Grupo Falabella needed to turn AI into a cross-functional business capability that improves customer experience, sales conversion, service responsiveness, and logistics efficiency across its five business units rather… US$265 million to technology in 2026 Banco Falabella app · Falabella.com · cameras 
NVIDIA Sample Jun 2, 2026 Revolut Revolut needed to replace siloed fraud, credit, marketing, and recommendation models that depended on months of feature engineering with a single behavioral foundation model that could learn from raw financial event… 16.1% improvement in credit risk prediction performance from NVIDIA Llama-Nemotron-Embed-1B-v2 NVIDIA H100 GPUs · NVIDIA Llama-Nemotron-Embed-1B-v2 · Nebius AI Cloud 