AWS Helps TVCMALL Build Generative AI Application in a Month, Cutting Translation Costs by 40%

Specializing in mobile accessories, TVCMALL has broadened its product categories to cater to client needs, now including consumer electronics, home & garden, outdoor sports, and automotive accessories. TVCMALL’s platform offers a unique experience by merging the strengths of wholesale suppliers with independent platform services, supporting multiple languages like English, French, German, Italian, Spanish, Portuguese, and Chinese.

For international markets, localizing product pages is vital for enhancing customer experience and broadening international reach. With over one million online SKUs and hundreds of new products listed weekly, TVCMALL demands multilingual translations of product information. Previously, this process relied on traditional translation methods and professional manual proofreading, leading to slower product launches and higher labor costs. These traditional services often offered literal translations that lacked cultural sensitivity, terminology accuracy, and language coverage, occasionally resulting in omissions and creating unprofessional impressions that necessitated additional proofreading. Moreover, traditional translation services or some LLM (Large Language Model) translations were costly and inefficient for translating large volumes of product descriptions. Additionally, sourcing information from multiple channels like suppliers, developers, and domestic wholesalers often led to unstandardized and unstructured original data, posing significant challenges for information extraction and content development.

To address these challenges, TVCMALL began partnering with AWS in 2024 to explore an AI-driven smart translation solution aimed at enhancing translation quality, improving customer satisfaction, standardizing data extraction, and reducing translation and customer communication costs.

Since 2015, TVCMALL has partnered with AWS, gradually migrating essential business systems like sales, ERP, and BI to AWS. For a decade, these systems have run smoothly, even during peak business periods, underlining TVCMALL’s zero tolerance for instability. As generative AI matures, AWS provides a holistic end-to-end AI technology stack that includes robust computing and storage resources, model-building tools based on Amazon Bedrock, multiple leading foundational models, and expert industry insights. Considering these resources, TVCMALL chose AWS as its partner to deeply assimilate generative AI technology within its operations.

To address TVCMALL's needs, AWS utilized services like Amazon Bedrock and Amazon Nova to integrate generative AI features across multiple applications on TVCMALL's global e-commerce platform. The scenarios encompass precise multilingual webpage translation, efficient aggregation and generation of product information, optimized image content processing and material re-drawing, providing solid technical support for TVCMALL’s global expansion and innovation.

Optimizing the Translation Workflow for Precise and Professional Multilingual Webpages

TVCMALL collaborated with AWS to implement Amazon Bedrock’s Anthropic Claude 3.5 and Amazon Nova LLM series, enabling real-time translation of new product entries and batch translation of millions of existing products. Whether working with factories or listing self-developed products, generative AI has helped TVCMALL organize and standardize varied foundational data into a consistent format, facilitating automatic multilingual page translations with professional quality and removing the need for manual proofreading.

During the development phase, AWS worked closely with TVCMALL to solve three key LLM application issues using Claude 3.5 Sonnet and Claude 3.5 Haiku through Amazon Bedrock. They tackled the LLM hallucination issue to prevent inaccurate translations or irrelevant information and ensure more accurate and reliable translation results; optimized e-commerce scenario prompts through customized prompt tuning for better adaptation to e-commerce contexts, ensuring accurate translation of e-commerce terminology to meet native e-commerce content creator standards; and ensured structured outputs with stable JSON translation results, efficiently managing a few unstructured outputs to maintain API call stability and efficiency with other business systems.

Notably, by replacing the previous translation solution with the Claude 3.5 model on Amazon Bedrock, TVCMALL substantially cut translation expenses, reducing real-time translation costs to one-third and existing product batch offline translation costs to one-sixth, saving thousands monthly. It not only expedited translation tasks but also reduced the number of Amazon EC2 instances required. TVCMALL plans to further refine models and processes to halve translation costs, leveraging the cost-effectiveness of the Amazon Nova model and efficient batch translation.

Enhancing Product Information Aggregation and Content Generation for Rich, Personalized Descriptions

For multimodal product information aggregation and listing scenarios, TVCMALL employs high-performance foundational models via Amazon Bedrock, integrating multimodal AI technology to automatically extract and consolidate information from varied sources into standardized webpage content. Operators can easily review and generate consistent product descriptions in either Chinese or English, which are then translated into local languages to streamline and speed up product listing processes.

For instance, information such as product highlights, descriptions, features, and specifications can be extracted from PDFs, Word documents, TXT files, scans, spreadsheets, images, and even audio records, transforming into a consistent format or new content, enhancing product descriptions with richness and personalization.

TVCMALL also plans to incorporate Amazon Nova for image content processing and product material re-drawing. During image and material generation, each product (like phone cases) requires multi-angled shots and color renderings. With hundreds of products and thousands of images launched on TVCMALL each week, this process consumes a significant amount of graphic design resources. The application of Amazon Nova enables efficient secondary processing, including texture enhancement and slogan integration, reducing editing time and optimizing content generation, allowing designers to focus on high-value tasks such as customer interaction.

In the fast-evolving landscape of generative AI, TVCMALL, in partnership with AWS, developed, tested, and launched the AI-driven smart translation solution within just one month. This ongoing optimization in translation quality has substantially enhanced customer experiences and accelerated product launches. Meanwhile, AWS’s robust infrastructure guarantees the smooth operation of TVCMALL’s application systems, delivering reliable and efficient B2B e-commerce services to users worldwide.

Enhanced Translation Quality Enhances Global Customer Experience

The models deployed via Amazon Bedrock have demonstrated excellent performance, delivering natural and precise multilingual experiences that earn commendations from customers and partners alike. Product descriptions, specifications, and all text content visible on webpages are tailored to suit local consumer preferences, enabling users to quickly and easily find the information they need, thereby increasing satisfaction levels.

Streamlined Product Listing Boosts Efficiency by 30%

Previously, product listings occurred once weekly, but with the advent of generative AI technology, the journey from market research to product listing has been significantly streamlined. By automating image processing and multilingual translations, TVCMALL can now complete product listings in just one to two days, boosting listing efficiency by 30%. In the past, three people were required to list 300 SKUs in a day; now, only two are needed, substantially reducing the time-to-market.

Translation Costs Slashed by 40%, Significantly Improving Content Production Efficiency

Leveraging Amazon Bedrock, TVCMALL has eliminated the need for dedicated translators and manual proofreading of traditional machine translations. Both real-time and batch translation costs have been cut by 40% overall. Generative AI’s capability to extract key information from multimodal data, such as images and documents, automates the conversion from raw product data to structured webpage content, significantly increasing operational efficiency by minimizing manual intervention.

While TVCMALL is optimistic about generative AI’s potential, uncertainties remain in certain applications and capability alignments. Through close collaboration with AWS, TVCMALL is transforming generative AI from a theoretical concept into tangible productivity gains, achieving cost reduction, efficiency enhancement, and quality improvement in suitable scenarios. Moving forward, with AWS support, TVCMALL plans to deepen its use of generative AI in customer relationship management (CRM), streamlining processes like order status queries using advanced model capabilities. Additionally, in development process restructuring, generative AI will assist in code writing, requirements validation, and testing to enhance software quality and project success rates.