ListingAgent: Gemini makes a bulk AI listing creator

Writing great product listings is a pain; even with AI you have to copy-paste a lot. ListingAgent makes a Gemini powered listing engine with you, from your needs. Your bespoke solution built to scale!

Comment

Inspiration

Creating listings can be a pain for resellers or ecommerce companies.

Lots of items in a messy spreadsheet, and each one needs a title, description, 13 tags, and pricing. You could paste them into a chatbot but that doesn't scale. The bottleneck isn't AI text generation, it's everything around it: messy data, brand voice, platform formats. Gemini 3 can handle all of that end-to-end.

What it does

ListingAgent takes product data in any format — CSV, JSON, pasted text, or images — and produces platform-ready listings through four steps:

  1. Discovery: Gemini uses code execution to analyze your data and write an extraction script, self-correcting on errors. For images, it uses vision.
  2. Interview: A short chat to learn your brand voice, platform, and requirements.
  3. Recipe building: Gemini builds a prompt template, output schema, and validation function with you — tests against samples, scores with code checks + LLM judge, and iterates until quality passes.
  4. Batch execution: The recipe runs across all products in parallel. You get upload-ready spreadsheets you can use for your chosen platform.

Schema fingerprinting means repeat uploads skip straight to step 4.

How we built it

Gemini does the heavy lifting, we just try to give it the best conditions. Runs on a $4/month droplet with a super light tech stack: FastAPI, vanilla JS, 11k lines across 12 files. Gemini 3 Flash can do the whole job - code execution, structured output, and guiding the user through the process.

Challenges we ran into

  • Evaluation: We needed a smart but not overbearing system to ensure quality outputs. We decomposed into binary pass/fail to avoid everything scoring "ok".
  • 512MB RAM: Wanted to keep it as light as possible, it works!
  • Messy data: Multiple extraction strategies (code execution, direct LLM, vision) to handle whatever arrives.

Accomplishments that we're proud of

  • Self-correcting code generation — Gemini writes Python, we run it, errors feed back, it fixes itself.
  • The recipe system — prompt, schema, and validation built with the user, iterated in real time.
  • Messy spreadsheet to 50+ polished listings in minutes, on a $4 server.

What we learned

A structured pipeline beats a fully agentic lobster sandbox on price at least. Also: more control, less risk. And Gemini writing extraction scripts on the spot beats pre-built parsers every time.

What's next for ListingAgent

Execution sandbox hardening, direct publishing to marketplace APIs, AI image enhancement, and multi-language generation.

Share this project:

Updates