Enterprise AI Document Translation Platform
Enterprise AI translation platform delivering 97% terminology accuracy (2M+ medical terms), 92% brand voice consistency, and 98% format preservation for leading pharmaceutical and medical device enterprises. Deployed for 3 major healthcare companies after competitive wins against incumbent vendors — featured as a keynote at AWS Summit China 2025, generating 10+ new business opportunities.
Translation Features
PDF: Academic Paper (English to Japanese)
User Experience Demo
Complete workflow demonstration: from document upload to terminology-enhanced translation output
Project Overview
An intelligent document translation platform built for enterprise pharmaceutical and healthcare clients with complex domain-specific terminology and strict regulatory style requirements. Started as a pilot engagement with an industry-leading pharmaceutical company and expanded to 3 major healthcare enterprises through competitive differentiation — winning against established translation vendors on terminology accuracy, brand voice consistency, and format preservation. The system handles medical device manuals, regulatory documents, and academic papers while maintaining pixel-perfect formatting across languages.
Core Technical Features
Architected terminology matching system using OpenSearch Percolator for 2M+ medical terms from WhoDrug database. Achieved 97% accuracy and reduced 500-page document translation from weeks to 30 minutes. Adaptive multi-scale matching: AC Automaton for small datasets, OpenSearch Percolator for medium-scale fuzzy matching, and enhanced OpenSearch with custom tokenization for pharmaceutical databases.
Engineered hybrid RAG architecture combining BM25 and vector embeddings with style transfer algorithms, achieving 92% brand voice consistency across pharmaceutical clients. Features historical translation corpus analysis, intelligent style pattern retrieval based on content similarity, and continuous learning from professional translator feedback.
Engineered spatial parsing algorithms handling Chinese-English text expansion (40% length variance) while maintaining 98% visual formatting accuracy for regulatory documents. Advanced PDF rendering with Font Shrink and Block Merge algorithms, plus XML-based Word document processing preserving table structures, font styles, and complex elements.
Led technical design for multi-model serving architecture supporting 6+ LLMs (Llama, Qwen, DeepSeek) on distributed GPU clusters using vLLM and LiteLLM. Built scalable FastAPI microservices with ALB, ECS Fargate containerization, and DynamoDB for state management, enabling seamless model switching and load balancing.
AWS CDK-based infrastructure as code for one-click deployment: Serverless architecture with Lambda functions, Step Functions orchestration, auto-scaling capabilities, and integrated monitoring. Complete system deployment from infrastructure provisioning to service configuration in under 20 minutes.
Translation Workflow
Translation workflow consists of offline and online phases: offline processing handles terminology databases and reference documents, online retrieval provides context for LLM translation based on similarity
System Architecture
Complete system architecture showing the end-to-end translation pipeline with AWS services integration and scalable processing components
Key Achievements
AWS Summit China 2025 — Featured Talk
Million Terms & Faithfulness, Expressiveness, Elegance: The Secret Behind Intelligent Document Translation
AWS Summit China 2025 · Featured speaker to 200+ attendees
Tech Stack
AI/ML & LLM Infrastructure
Search & Data
Languages
Backend & API
Frontend
Document Processing
Infrastructure & Deployment
All data, metrics, screenshots, and demos shown are based on publicly available information or anonymized data.