AGENTIC.LIVE — The broadcast booth in a browser tab
Link to open source: https://github.com/Arun07AK/agentic-live
Link to Live Project: https://agentic-live.vercel.app
AGENTIC.LIVE — The broadcast booth in a browser tab
Point your phone or laptop camera at any cricket screen — TV, projector, live stadium — and six Gemini agents become a full broadcast booth in real time.
Six Gemini agents working together:
- VISION watches video frames and generates Hinglish commentary in your chosen persona's voice (Bhogle / Shastri / Sidhu).
- STAT OCRs the on-screen scoreboard so the system can't hallucinate scores.
- DIRECTOR gates every line — APPROVE / REWRITE / HOLD — preventing stale or off-key output.
- CROWD reads viewer emoji reactions in 5-second windows and tells the Director when the room is on fire.
- TRANSLATOR fans out captions to English, Hindi, Tamil and Bengali — each viewer phone receives their preferred language with sub-second latency.
- ISL LISTENER independently listens to broadcast audio via the microphone, transcribes via Gemini, generates Indian Sign Language gloss, and feeds it into a 3D signing avatar.
Problems solved:
- India's 1.8 crore Deaf cricket fans get ISL on EVERY match — not just the 10 marquee broadcasts a year that have human interpreters.
- Multilingual viewers in the same watch-party room get captions in their language without separate streams.
- Anyone with a phone becomes their own broadcast booth — gully cricket, school finals, college tournaments, the IPL on your TV.
Built solo in 4 hours at the Build With AI :: Agentic Premier League hackathon, GDG Cloud New Delhi, 03 May 2026.
Tech stack:
AI / Google: Gemini 2.5 Flash via @google/genai SDK powering all six agents · Audio + Vision multimodal input (mic chunks + JPEG frames inlineData) · Gemini API Paid Tier (project-billed) · AntiGravity IDE (Google's agentic IDE).
Frontend: Next.js 16 (App Router, React 19, TypeScript strict) · Tailwind CSS v4 + shadcn/ui · Inter Tight + Geist Mono · Browser SpeechSynthesis API for persona-mapped TTS.
Realtime: PartyKit on Cloudflare Workers (sub-second pub/sub) · partysocket React hook for bidirectional commentary fanout + emoji reactions back to the caster.
Accessibility: Vaani 3D ISL signing avatar integrated via /embed postMessage protocol — Audio → Gemini transcribe → ISL gloss → postMessage → avatar signs.
Hosting: Vercel (Next.js production) · Cloudflare Workers (PartyKit room server).
Open-source code lifted: google-gemini/live-api-web-console (Apache-2.0) — audio worklets + GenAI client patterns.
Links:
