LingoStream

At a Glance

The Challenge

Live audio content, whether a stream, a broadcast, or a public address, is built around a single spoken language. Existing translation tools bolt subtitles onto that single stream after the fact: flat, delayed, and one-size-fits-all for every viewer regardless of what they actually speak.

Two problems compound this. First, translation is usually literal. Excitement, urgency, and humour get flattened into neutral text, and domain-specific language, slang, jargon, technical terms, often gets mistranslated or dropped entirely. Second, the language decision sits with the broadcaster, not the audience. A streamer or speaker can only address one language at a time, which quietly excludes every viewer who doesn't share it.

The Inspiration

LingoStream began as preparation for a run of AI and cloud engineering events, including an Agentic AI Demo Day. Rather than arrive with a finished, polished product, the goal was to design and build something live, in the room, as a demonstration of applied agentic engineering.

The starting idea was narrow: a translator for game streamers, built to handle fast-moving domain slang like "clutch" and "respawn" without losing meaning. Two things pushed it further. The first was a simple observation about how language actually works in a live audience: nobody watches or listens as a group that shares one preferred language. Each viewer should be able to choose their own, independently, without the broadcaster having to do anything differently. The second was a market reality closer to home: Africa alone carries more than 2,000 languages, and mainstream translation infrastructure treats most of them as an afterthought. Testing against Amazon Translate surfaced a concrete gap early on: isiZulu and isiXhosa, two of the most widely spoken languages in South Africa, are not supported. That gap became a design input rather than a footnote.

From there, the concept widened from gaming into a general-purpose translation layer for any live spoken content: education, healthcare, government communication, corporate broadcasts, religious services, sport, and financial services, each with their own tone and terminology to preserve.

The Approach

LingoStream was designed around one architectural principle: the streamer's audio is the only input. Every other decision, which language a viewer hears, how urgency or excitement comes through, which domain vocabulary applies, is handled by the platform, not the broadcaster.

The result is a viewer-controlled model rather than a broadcaster-controlled one. Each viewer selects their own output language independently, and the system carries the speaker's emotional register (calm, excited, urgent, instructional) through the translation rather than producing a neutral text-to-text conversion.

Technical Architecture

LingoStream is a serverless, event-driven, multi-agent platform built on AWS, following a pipeline pattern:

Streamer audio (mic / OBS / stream source)
        |
        v
Amazon Transcribe Streaming  (real-time, multilingual detection, 200-500ms chunks)
        |
        v
EventBridge Pipe  (routes transcript chunks to the agent runtime)
        |
        v
Orchestrator Agent  (Strands Agents SDK + Bedrock AgentCore Runtime)
        |
        |- Language Detection Agent   (confirms language, handles code-switching)
        |- Translation Agent          (Amazon Translate, per active viewer language)
        |- Tone-Matching Agent        (Claude via Amazon Bedrock, energy and register preservation)
        |- Vocabulary Agent           (Bedrock Knowledge Base, domain terminology)
        |- Delivery Agent             (API Gateway WebSocket, per-language channels)
                                            |
                                       DynamoDB
                              (viewer registry, session state,
                               vocabulary preferences, audit log)

Each agent has a narrowly scoped, structured-output system prompt. The Orchestrator classifies every chunk by language, energy level (calm, conversational, excited, urgent, instructional), and domain (gaming, medical, legal, financial), then routes it accordingly. The Tone-Matching Agent is deliberately constrained: it may not change meaning or add words the speaker did not say, only carry emotional register into the target language.

A cost- and latency-conscious design choice underpins the pipeline: the full agentic path (tone-matching, vocabulary lookup) is invoked selectively, on roughly 30 to 40 percent of chunks, based on detected energy level and domain relevance. Lower-signal chunks pass through base Amazon Translate directly. This keeps the system responsive at live-streaming speed while reserving the heavier reasoning for moments where tone actually matters.

Google AI Studio and the Google ADK were used during development as a testing environment for prompt engineering and agent behaviour validation, separate from the production AWS backend.

Status

LingoStream exists today as a working proof of concept, architected and built as a live demonstration of multi-agent orchestration on AWS. It shows a full pipeline, from raw streaming audio through language detection, tone-aware translation, and per-viewer delivery, running end to end.

A note on this section: specific reception or follow-up outcomes from Demo Day aren't in the record used to draft this case study. If there's a concrete result worth including (audience response, interest generated, a next step it led to), it's worth adding here before this goes live, since that's usually the section a reader trusts most.

What This Demonstrates

• Design and delivery of a production-shaped, multi-agent architecture on AWS (Transcribe, Translate, Bedrock, Strands Agents SDK, AgentCore Runtime) under real event time pressure
• Systems thinking that starts from a real, underserved gap (African language support in mainstream translation tooling) rather than a generic feature list
• Cost-aware agentic design: selective intelligence invocation rather than running every chunk through the most expensive reasoning path
• A broadcaster to platform to viewer control model that generalises cleanly beyond its original use case

LingoStream is one example of the applied AI and cloud architecture work behind ByTheodore's AI implementation and cloud engineering services.