Amazon needed to correct misspelled product-search queries in real time so shoppers could find products quickly without sacrificing latency, throughput, or the time required to tune inference configurations for production.
Reported impact
Amazon achieved real-time spell correction with under-50-millisecond latency and 5x throughput for the T5 model, while Triton Model Analyzer reduced configuration work from weeks to hours so teams could improve search quality faster.
Amazon implemented AI-driven spell correction for product search by deploying the T5 natural language processing model on GPUs in Amazon Web Services and optimizing the serving stack with NVIDIA inference software. Search-bar queries are routed into a T5 text-to-text model, while NVIDIA Triton Inference Server provides the production serving layer and NVIDIA TensorRT accelerates deep-learning inference so corrections can be returned fast enough for interactive shopping. Amazon also used Triton Model Analyzer to tune and validate the best inference configuration instead of spending weeks on manual experimentation. The result is a production architecture where the model, serving runtime, and tuning tools work together to keep search responsive while scaling higher query throughput for a global retail workload.
A shopper's search query enters Amazon's search bar, the T5 model generates corrected text on AWS GPUs, TensorRT and Triton handle optimized low-latency inference, and the corrected query is returned to the retail search experience before results are displayed.
Amazon improves the customer experience with ai-driven, real-time spell check for product search.
The NVIDIA solutions respectively delivered under 50ms of inference latency and 5X the throughput for the T5 model, using NVIDIA GPUs on Amazon Web Services (AWS).
The Triton Model Analyzer also reduced the time needed to find optimal inference configuration from weeks to hours.
To accelerate text correction, they leverage NVIDIA AI inference software, including NVIDIA Triton™ Inference Server, and NVIDIA® TensorRT™.
To achieve the desired results, Amazon uses the Text-To-Text Transfer Transformer (T5) natural language processing (NLP) model for spelling correction.
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
NVIDIA Sample Aug 14, 2026 Shopee 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… 113x growth from 3 billion to 340 billion monthly tokens Compass v3.5 · NVIDIA A100 GPUs · NVIDIA H100 GPUs NVIDIA Sample Jul 24, 2026 Instacart Instacart needed real-time visibility into in-store shopping behavior, shelf conditions, and basket changes so retailers could reduce shrink, improve recommendations, and personalize physical-store experiences that… 3x year-over-year Caper Cart growth NVIDIA Dynamo · NVIDIA Hopper GPUs · NVIDIA Jetson Orin NX modules NVIDIA Sample Feb 2, 2026 Fisia Group Fisia needed to improve fit confidence, conversion, and basket size for Nike shoppers while reducing returns and making the same personalized experience work consistently in stores, on mobile devices, and online. 128% ROI Doris.ai · NVIDIA A10 Tensor Core GPUs · NVIDIA TensorRT 