Most large language models can speak multiple languages. That doesn't mean your AI application can.

In this episode, Dex and Vaibhav break down one of the most overlooked challenges in AI engineering: building agents that reliably work across multiple languages. While multilingual support seems straightforward on the surface, many production AI systems accidentally steer users back toward English because their prompts, tools, business logic, and agent workflows are all designed around English-first assumptions.

Key Takeaways
• Why multilingual support is harder than simply choosing a multilingual model
• How English-heavy prompts can unintentionally influence agent responses
• The tradeoffs between separate language pipelines and translation-based architectures
• Why reliability matters more than average performance when serving real users
• How voice agents solve a similar problem using normalization layers
• Building multilingual systems that are easier to evaluate, maintain, and improve
• When it makes sense to use translation layers instead of duplicating your entire pipeline
• How to think about latency, cost, and user experience in global AI applications

Summary
Dex and Vaibhav explore the architecture patterns behind multilingual AI applications and explain why many seemingly capable agents struggle when users interact in languages other than English. The discussion begins with a simple question: if modern models already understand dozens of languages, why do multilingual applications still fail?

The answer comes down to how AI systems are built. While general-purpose assistants rely heavily on user input, production AI applications often contain large amounts of prompts, instructions, tool definitions, and business logic written in a single language. As more responsibility shifts from the user to the application, those English-first assumptions can begin influencing outputs in unexpected ways.

The episode compares different approaches to multilingual support, including dedicated language-specific pipelines and translation-based architectures that normalize user input into a common language before processing. Along the way, Dex and Vaibhav discuss eval strategies, translation quality, latency tradeoffs, voice agent design, intent extraction, and practical techniques for building AI systems that remain reliable across global audiences without multiplying maintenance costs.

TIMESTAMPS
00:00 Teaser
00:30 Welcome to the Show
01:11 Building AI Agents for a Global Audience
03:11 Don't LLMs Already Speak Multiple Languages?
06:10 The Hidden Burden of AI Application Design
07:18 Why Your Agent Keeps Replying in English
08:41 Two Ways to Build Multilingual AI Applications
11:48 The Difference Between 99% and 99.99% Reliability
13:03 Why Dedicated Language Pipelines Don't Scale
14:43 One Pipeline, Every Language
18:13 What Happens When Users Mix Languages?
22:49 Building a Real Multilingual Travel Agent
26:29 How Do You Test AI in Languages You Don't Speak?
31:06 Why Most AI Tools Still Assume English
33:19 Building a Multilingual Travel Agent Live
35:46 The Translation Pipeline in Practice
38:02 Simple Optimizations That Reduce Latency
42:17 Should English Users Skip the Translation Layer?
44:00 How Output Schemas Accidentally Influence Models
45:15 Building Internal Backpressure With Better Testing
48:41 Putting the Multilingual Pipeline to the Test
50:00 How Much Should You Trust the Model?

#AIThatWorks #AIAgents #AIEngineering #MultilingualAI #PromptEngineering #ContextEngineering #LLMs #SoftwareEngineering #DeveloperTools #ArtificialIntelligence