Multimodal Listing Pipeline
Photo → recognition (vision) → bilingual listing card (LLM) → price calibration against live comparables → human review → publish. 110 tests.
Architect & solo developer
solo · solo
- Python
- Gemini Vision
- DeepSeek
- HITL
Expand the deep-dive ↓
What doesn't work
Building a product listing from a photo by hand is slow: recognize the item, write a description in two languages, set a realistic price, format it for the marketplace.
Architectural approach
A multimodal pipeline: a vision model recognizes the item from a photo → an LLM writes a bilingual card → a pricing step calibrates against live comparables → a human confirms before publish. Human-in-the-loop by design.
Architect & solo developer
Designed and built the multimodal pipeline: a vision model recognizes the item from a photo, an LLM drafts a bilingual listing card (title/description), a pricing step calibrates against live comparables, a human confirms before publish. 110 tests, live-verified.