Modernizing a Multilingual Meeting Platform

The Challenge

Our client operates a multilingual meeting platform that enables participants to converse in their own language — breaking down communication barriers with live simultaneous human interpretation. The platform serves organizations that need real-time multilingual communication: international conferences, global business meetings, and cross-border collaboration.

The product had strong market traction, but the technology behind it was struggling to keep up. The existing codebase was built on a legacy stack that made it increasingly difficult to ship new features. Performance issues were creeping in as the user base grew. The mobile apps — critical for on-the-go meeting participants — needed to be completely rebuilt. And the team had ambitious plans that the current architecture couldn't support: integrating with third-party meeting platforms and adding AI-powered translations alongside the existing human interpretation workflow.

They needed engineers who could embed into their existing team, understand the product deeply, and move fast without breaking what already worked.

Our Solution

We joined the client's team through a staff augmentation engagement — two Jyaasa engineers working as an integrated part of their development team. Same standups, same tools, same codebase. The engagement has been running for over two years and continues today.

Rebuilding mobile from the ground up

The most immediate priority was the mobile experience. We rebuilt both the iOS and Android apps from scratch — native Swift for iOS, native Kotlin for Android. Going native (rather than cross-platform) was the right call here: real-time audio streaming, low-latency interpretation feeds, and platform-specific UX patterns all demanded deep access to device capabilities. The result was a faster, more reliable mobile experience that matched the quality users expect from a communication tool they rely on daily.

Modernizing the web platform

On the web side, we worked within the existing Angular codebase to systematically address tech debt and performance bottlenecks. Rather than proposing a full rewrite — which would have stalled feature development for months — we took an incremental approach: refactoring critical paths, improving load times, and restructuring the architecture to support the features on the roadmap. The backend runs on Node.js with a Python FastAPI service layer for AI workloads.

Launching AI-powered translations

The most impactful addition was extending the platform to support real-time AI translations alongside the existing human interpretation model. This meant building a new service layer that could process audio streams, run them through AI translation models, and deliver translated output back to participants — all with latency low enough to feel like a natural conversation. The AI translation capability opened the platform to use cases where human interpreters aren't available or cost-effective, significantly expanding the product's addressable market.

Third-party integrations

We also built integrations with major third-party meeting platforms, allowing the multilingual capabilities to plug into tools teams already use. This required working with external APIs, handling authentication flows, and ensuring the interpretation and translation layers worked seamlessly within meetings hosted on other platforms.

Key Technical Decisions

Native mobile over cross-platform. For a real-time communication product handling live audio streams, native development (Swift and Kotlin) gave us the low-latency performance and platform-specific audio APIs that cross-platform frameworks couldn't match.

Incremental modernization over full rewrite. The web platform was a running product with active users. We modernized it piece by piece — improving architecture and performance without freezing feature delivery.

Dedicated AI service layer. AI translation runs on a separate Python FastAPI service rather than being embedded in the main Node.js backend. This keeps the translation workload isolated, independently scalable, and easier to iterate on as models improve.