LiveGSeller: AI eBay listing agent.

LiveGSeller turns your camera, voice, and photos into ready-to-publish eBay listings.

Comment

Problem

Selling things online sounds easy.
Actually creating listings is not.

Titles, descriptions, categories, shipping, pricing, photos – every marketplace UI adds more tiny decisions and clicks. I personally kept procrastinating on selling my own stuff just because I hate filling out those forms.

Inspiration

The idea came from a very simple, very real pain:
yesterday I was thinking about how many things I want to sell,
and how much I avoid it because of the UI and forms.

I searched for an AI agent that could watch the screen and fill listings for me.
I found tools that generate text, but nothing that really drives the on‑screen workflow end‑to‑end. So, one day before the deadline, I decided to build the agent I wanted for myself.

By background I’m a psychology student, not a professional software engineer.
But I’ve been learning to code through AI agents and hackathons, so this project is also me testing how far I can go in ~24 hours with Gemini Live.

What is LiveGSeller?

LiveGSeller is a multimodal AI agent powered by Gemini Live that helps create product listings from your voice while understanding the live UI on the screen.

You talk, the agent:

  • sees the current marketplace listing page
  • listens to your voice description of the item
  • suggests titles, descriptions and key fields
  • and then fills in the form step by step in real time.

Instead of manually fighting each input field, you collaborate with a screen‑aware agent that understands both your intent and the current UI state.

How it works (high level)

  1. The browser captures the screen and streams visual context into a Gemini Live session.
  2. Your voice is captured via Web Audio / MediaDevices APIs and streamed as text/audio into the same session.
  3. The agent builds an internal view of the current UI (what fields exist, what is missing) and of the item you want to sell.
  4. Based on that, it decides which field to fill next and how to interact with the page (typing, clicking, choosing options).

Conceptually, you can think of it as optimizing a simple objective:

[ \text{maximize}(\text{listing_quality} - \text{user_effort}) ]

where listing_quality is about clarity, completeness and relevance of the final listing,
and user_effort is how many manual steps you still have to perform.

How I built it

  • Frontend: React + TypeScript for the UI, screen capture, and audio capture.
  • Backend: Node.js + Express as a thin API layer orchestrating Gemini Live sessions.
  • AI layer: Google Gemini Live API via the Google GenAI SDK, keeping a live multimodal conversation that mixes UI screenshots and user speech.
  • Cloud: Deployed on Google Cloud Run, with Google Cloud Firestore / Storage to persist basic session and config data.

All of this was hacked together in roughly one day: from “I wish this existed” to “I can record a demo of it actually driving a real listing UI”.

Challenges

Building a live, screen‑aware agent in 24 hours came with some fun challenges:

  • Figuring out how to stream just enough visual context to Gemini Live to keep latency low but still give the model the full UI state.
  • Designing prompts and tool calls so the agent doesn’t just describe the page, but actually takes actions in a reliable order.
  • Handling edge cases in the listing flow (missing required fields, validation errors, unexpected UI states).

The biggest personal challenge: doing all this as someone who doesn’t come from a classic CS background, and learning as I go. This project is both a real tool I wanted for myself and a proof that with Gemini Live you can ship a useful, screen‑aware agent in a weekend.

Built With

Share this project:

Updates