~/work / quickposters
QuickPosters
Generative AI posters for Indian small businesses. A retrieval step personalizes each poster, image models generate it, and a VLM judge checks it before the user sees it.
- text-to-image
- semantic-retrieval
- vlm-as-judge
01 / problem
Why it needed to exist
Shopkeepers, resellers and freelancers need fresh festival, offer and greeting posters every day, in their own language, but can’t design and can’t pay a designer.
02 / what i built
The approach
01
Preference-aware retrieval
Semantic search over vector representations of styles and content, conditioned on the business profile, so generations match the shop rather than a generic template.
02
Two image models
Imagen 4 and FLUX handle generation inside the same pipeline.
03
VLM as judge
A multimodal model scores every generated image for visual quality, alignment with the message and content suitability, as part of an agentic pipeline, before it reaches the user.
04
Built for Bharat
Logo and shop photo are auto-inserted, and posters can be made in English, Hindi and Marathi with OTP login and no passwords.
03 / architecture
How data flows
- 00Business profilecategory · logo · language
- 01Semantic retrievalembeddings → relevant styles & context
- 02GenerateImagen 4 / FLUX
- 03VLM judgequality · message alignment · suitability
- 04Composelogo, text, language
- 05Sharedownload & share, ad-free on premium
04 / results
What it did in the real world
- 1K+ downloads on Google Play
- 5.0★ average from 11 verified reviews
- Subscription with in-app purchases