Inspiration
Billions of people are locked out of useful AI by language. AgentAsia treats multilingual access as a first-class product feature, not an English app with a translation bolt-on.
What it does
AgentAsia is a browser-native AI platform that forms adaptive agent teams to research, write, translate, and act in the user's language.
- Pick a language (Bengali, Urdu, English, Arabic, German, and more); UI, agent replies, and voice stay in that language
- NVIDIA Nemotron 3 Nano 30B A3B on Nebius Token Factory for chat and reasoning (live sub-second latency ~0.9s)
- Tavily runtime web search with citations: the model refuses to invent live prices; it searches and sources instead
- Voice input/output with honest language coverage (no silent English fallback)
- Multi-agent orchestration: chat / code / agent lanes form teams automatically
- Server-side spend-guard so credits cannot run away
How we built it
- NVIDIA Nemotron (Nano 30B A3B / Super / Ultra lanes) via Nebius Token Factory OpenAI-compatible inference API
- Tavily search API for live facts and citations
- TypeScript, React, Vite browser app with multi-agent task orchestration
- Voice stack with verified Asian language coverage tables
- Circuit-breaker spend guard metering every Token Factory call
Challenges we ran into
Language coverage is uneven (some languages have no verified TTS). We made that honest: speak where verified, text plus a coverage badge where not, and never a silent English substitute.
Accomplishments that we're proud of
- Live demo already running on Nemotron via Nebius Token Factory
- Measured real costs (~$0.000002/turn Nano) and built a spend guard
- Tavily is load-bearing: when the model cannot know live facts, it searches and cites
What's next for AgentAsia
More languages, deeper agent skills, and always-on private assistant workflows.
Demo: https://agentasia.vercel.app/ Track: Best Apps and Agents License: MIT
Built With
- multi-agent-orchestration
- nebius-token-factory
- nvidia-nemotron
- react
- tavily
- typescript
- vite
- voice/tts
Log in or sign up for Devpost to join the conversation.