A seller’s photos shouldn’t leave her laptop. ListingForge writes the listing without them ever going online.

*What I Built
*
Large batches (more than 10 items) need a license at ₹499 a month. A seller who lists 60 new products a month at roughly 40 minutes each spends about 40 hours writing listings. ListingForge turns that into a batch run she starts and walks away from, so one month’s license pays for itself in the first hour she gets back.

Code

GitHub: https://github.com/Jha-2022/ListingForge.git

How I Built It

Real example of a limit the model broke:
On my first Amazon India test, the model wrote a 163-character Etsy-style title for a 140-character Etsy field and 14 tags instead of 13. One tag, handmade-soy-candle-lavender, was 28 characters. The rules engine trimmed the title at the last whole word under 140, dropped the extra tag, and rejected the over-long one, so the seller never saw an invalid listing.

Hindi, English and Hinglish note:
Hinglish needed the most prompt work. Without examples the model either went into formal Hindi or into stiff, translated English. Adding two few-shot lines fixed it.

• Before: “Yeh candle bahut hi sundar aur prakritik hai.”
• After: “Ghar ko banaye cozy: handmade soy candle, lavender ki shaandaar khushboo ke saath. Diwali gifting ke liye perfect!”

What it got wrong

• Hallucinated materials. The model called a painted MDF box “hand-carved teak”. I added a prompt rule to describe only what is visible, plus an editable materials field the seller confirms before the listing is finalized.

• Speed on CPU. A single listing took a median of about 75 seconds with no GPU. That’s fine for a batch she starts and leaves, which is why the SSE progress stream exists, but it’s slow for one-off tweaks.

• Hinglish consistency. It sometimes drifted into pure Hindi in the third bullet. I fixed it by repeating the “Hinglish, Roman script” instruction at the end of the writer prompt.

• The byte limit with Hindi. Amazon’s 250-byte backend field counts UTF-8 bytes, and each Devanagari character takes 3 bytes. A Hindi search-terms line that looked short was over the limit. The rules engine now counts bytes, not characters.

Then I tested it on real products

I ran 12 real products from my cousin’s Jaipur block-print shop. She edited 5 of the 12 titles, mostly to add her brand name, and kept 9 of the 12 bullet sets unchanged. She rewrote the “story” paragraph for 3 products to sound more like her. That told me the brand voice profile needed more example sentences, so I raised the recommended number of samples from 2 to 4.

Why Does Open Innovation Matter?

Closed-model comparison:
A hosted frontier model would likely write a listing in about 5 seconds, against about 75 seconds on my CPU laptop, and its photo understanding is probably a bit sharper. But a seller with 200 products and constant regenerations would pay per call, and her unreleased photos would sit on someone else’s server. For this product, that trade-off is the reason it exists

My First User

The first real user was my cousin, who sells block-printed kurtas on Meesho from home. She had 12 products waiting to be listed for the Diwali season. ListingForge made 12 Meesho listings in about 15 minutes while she packed orders. She said the Hinglish titles sounded like her own, but that she wanted the materials field to be mandatory. That’s exactly why the “confirm materials” step exists. The bugs her use exposed were a ZIP with a space in a folder name failing to import, and the batch progress bar freezing on the last product. Both are fixed.

Prize Categories

• Best Use of Gemma: include this only if you tested it. Sample wording if you did: “I swapped the model to gemma3:12b using the VLM_MODEL variable and ran the same 30 products. Gemma took X seconds per listing and had Y hallucinated attributes versus Z for Qwen.”

• Local-first / open-weight: “Vision and writing both run on open weights on the seller’s machine, with no per-listing cost and no photos uploaded.”

• Other category: only add one if you actually entered it.

Checklist, answered

1. Eval: run the same 30 products through all three marketplaces and record the table above.
2. “Got wrong” story: use whichever of the four items actually happened to you.
3. User feedback: even one seller’s quote counts.
4. Prize categories: keep only the ones you tested.
5. Screenshots: Listing Studio, batch progress, marketplace switch.