VoiceTrade AI

VoiceTrade AI (Mira): a voice-first AI trading companion for India's next generation of retail investors

What it is

VoiceTrade AI is a voice-first stock market and paper trading app built for the Agora Voice AI Hackathon. Instead of navigating charts and menus, users talk to Mira, a real-time conversational AI, in English or Hindi, the way they would talk to a knowledgeable friend. Ask "how is Reliance doing," "what are today's biggest gainers," "how is the IT sector doing," or "buy 5 shares of Infosys," and get a natural, spoken answer grounded in live market data.

Why we built it

Stock market investing in India is still gatekept by English-heavy, chart-dense apps that assume a level of financial literacy most first-time investors do not have. Millions of Indians are comfortable talking about money in Hindi or Hinglish, but every trading app forces them into a foreign visual and linguistic interface. The result is that people either stay out of the market entirely, or invest blind. We wanted voice to be the actual product, not a bolted-on feature, so beginners could learn to trade by having a conversation.

How it works

• Agora's Conversational AI Engine handles the live voice connection and turn-taking.
• Sarvam converts speech to text and back, enabling natural Hindi, English, and Hinglish conversation.
• An agentic LLM backend (FastAPI, deployed on Fly.io) does the reasoning: it decides what the user needs and calls live tools instead of guessing.
• Every price, company fact, and portfolio figure comes from a real data call in that turn: live quotes and fundamentals from Yahoo Finance, NSE gainers and losers, IPO calendars, and sector indices, never invented.
• A full paper trading engine lets beginners practice with real prices and real portfolio mechanics without risking real money.
• The Android app (Kotlin, Jetpack Compose) is a clean visual companion to the conversation: cards for quotes, order previews, and disambiguation choices, so voice and screen work together.

What makes it different

• Order confirmation has a code-level safety gate independent of the LLM, so a misheard "yes" or a prompt injection attempt cannot place a trade.
• Tool-calling is grounded, so the assistant cannot assert stale or hallucinated facts about a company.
• The conversational feel itself, filler pacing, handling interruptions, and language switching, was tuned against real usage, not just a demo script.

Who it is for

Someone who has never opened a trading app before, and needs investing to feel like a conversation, not a console.