Building an AI service without training – AI translations with LLMs
summary
Translates short product titles into 13 languages by combining OpenAI embeddings and GPT completions with a vector search layer (FAISS initially, then OpenSearch + embeddings), pulling product metadata from Oracle and returning JSON-ready translations into the PIM and a Streamlit review app. Prototype uses FastAPI, embeddings for nearest-neighbour examples, and human-in-the-loop evaluation (BLEU and manual review); it reduces translator work but exposes risks: model drift, no SLA on hosted models, and costly human evaluation, though switching models is trivial.
AI-written summary · about
chapters (10)
- 0:00AI translation with LLMs
- 2:41Avoiding standard tools
- 5:32Fusion prompting
- 7:58Text embeddings for context
- 10:25Tech stack overview
- 11:51Integration with data
- 12:54Evaluating translation quality
- 14:39Model drift and API reliability
- 16:06API-based LLM advantages
- 17:54Summary
related talks
13mCustom Training of AI for Language Services: Translation is Just the Beginning
28mHow to Compete with OpenAI and Google
18mAI translation + AI agents = i18n made easy
40mDeploying Your Own Language Models: A Use Case Demo with Open Source Tools
27mBuilding AI applications: Successes and Lessons
39mExperiments with OpenSearch and AI
24mGenerative AI: Business or Feature?
45mCut Translation Costs by 50× for Millions of Product Sheets…