Deeptalk — Real-Time LLM Translation for Multilingual Meetings

Role: AI Engineer / Fullstack Engineer

Tech Stack: Go (Backend/API) · Next.js (Web) · Flutter (Mobile) · Node.js + Socket.IO (WebSocket) · LLMs (Speech Recognition & Translation)

Overview

Deeptalk is a production AI platform that delivers sub-second multilingual translation for live meetings. Hosts speak in their native language; participants receive transcribed and translated text in real time — covering global languages (English, French, Indonesian) and Indonesian regional languages (Javanese, Sundanese, Minang).

The system was designed to handle concurrent meeting rooms, low-latency broadcast, and the latency/cost trade-offs of running LLMs in a live conversational loop.

What I Built

• Go backend (Gin): REST APIs, auth, meeting-room orchestration, and connection management between web/mobile clients and the translation pipeline.
• WebSocket layer (Node.js + Socket.IO): Bidirectional low-latency broadcast so every participant receives the translation feed within milliseconds of the host speaking.
• LLM translation pipeline: Speech captured → transcribed → translated to each participant’s selected language → streamed back over WebSocket. Tuned prompts and provider routing to balance accuracy vs. latency vs. cost.
• Next.js web client for hosts and participants (SEO-friendly landing, fast room join flow).
• Flutter mobile client sharing the same realtime contract as the web app.

Engineering Highlights

• ⚡ Real-time architecture: WebSocket broadcast with per-room fan-out — translation results delivered in real time to every participant.
• 🧠 LLM-driven translation: Combined speech recognition and translation through LLMs, with prompt design tuned per language pair.
• 🌐 Regional language support: Modeling for Indonesian regional languages (Javanese, Sundanese, Minang) that off-the-shelf MT engines handle poorly.
• 🔁 Multi-platform parity: Single backend contract powering web (Next.js), mobile (Flutter), and partner clients.
• 🎯 Inclusive UX: Hosts and participants can switch language preferences mid-meeting without disrupting the conversation.

Why It Matters

Built for global teams, classrooms, webinars, and communities where accessibility and inclusivity in communication directly determine whether participants can take part. A real-world production system at the intersection of backend engineering and applied AI.