Techint
RAG assistant over the institutional memory of one of Latin America's largest industrial groups, deployed to production in 2023, when almost nobody knew how to do this yet.
Impact
20,000+
Articles indexed
Multilingual corporate content
80
Global communicators
Assistant users worldwide
15%
Search time reduction
Per query over internal archive
2023
Year in production
Before this was an established practice
Where they were
Techint is one of the largest industrial groups in Latin America: gas pipelines at 4,900 meters, LNG plants in Europe, operations in more than 13 countries. Their Communications team manages an intranet with more than 20,000 articles in multiple languages, which constitutes decades of institutional memory. Their writing team consulted this information to create content, verify journalistic coherence, and maintain brand consistency.
It was 2023 and language models had just hit the market. Almost nobody knew how to bring them to production inside a real company, and much less inside a multinational with 100 years of history to protect. At Techint, they already knew AI could change how the team works with that archive, but from the start they knew this meant doing it right.
What we found
The request was “a chatbot that answers questions about internal news,” but the challenge was even more important at that time: information retrieval over a massive, multilingual archive, temporal relevance (where recent content weighs more than historical), specific editorial requirements, and the need to cite sources with exact links. And all of this had to be deployed under the strictest security and compliance standards.
What we built
A conversational assistant with RAG architecture over more than 20,000 corporate articles. Vector database for semantic search, LLM for generation, frontend embedded via iframe in the intranet. All built within Techint’s security constraints.
Designed for the editorial workflow: answers with links to the original source, temporal awareness to distinguish recent from historical content, response length options. Feedback system (thumbs up/down) for continuous improvement. The pilot started with public content from tenaris.com and expanded to the full internal archive. deployr took ownership end-to-end: information architecture, embeddings, deployment, integration with the intranet.
What they can do now
Communicators across the world can now query decades of institutional history in natural language. What used to mean manually searching through a 20,000-article archive now has answers in seconds, with the source cited.
Why it matters
First generative AI project in production inside Techint, with all the scrutiny that implies in a multinational. Built in 2023, we put cutting-edge technology into practice to accelerate time-to-value in one of the world’s largest companies. How to make retrieval that actually works, how to operate within a multinational’s constraints, that an LLM cites well and doesn’t hallucinate in an editorial context: all of this is what we were able to provide to improve the efficiency of a creative and operational process.
“ They speak your language, understand your needs, and hold themselves to the standard we demand in a multinational. A stable team that goes beyond expectations.”
• Sebastián Flexas, Technology & Digital Manager – Communications, Techint