Instax
Semantic photo search over CLIP embeddings with pgvector, plus a written plan for turning the prototype into a product.
- Position
- Sole developer; prototype and product strategy
- Period
- August 2026
- Stack
- Python
- open_clip (CLIP ViT-B/32)
- PyTorch
- PostgreSQL
- pgvector
- ChromaDB
- FastAPI
- Alembic
- Docker
- Next.js
- Status
- Working prototype; code private (see below)
- Repo
- private (see below)
The problem
Someone standing at a location wants to know what a good photo here looks like and how to take it. That means finding reference images by mood, by an example image, or by place, and not by hashtag.
What works
An end-to-end retrieval loop: CLIP ViT-B/32 embeds every image to a 512-dimension vector; pgvector (HNSW, cosine) stores and searches them; a FastAPI server exposes /search/text, /search/image, and /search/location, with WebP thumbnails generated on the fly. The embedding job is batched in FP16 and resumable. The data layer was migrated from SQLite to Postgres with Alembic migrations.
The strategy document
It opens with what is wrong with the prototype: single-node with hardcoded paths; “location search” is a text embedding of a place name with no coordinates behind it; the photo tips are fourteen hardcoded dictionaries; two modules read from different tables; and a 2021 CLIP model cannot tell a golden-hour backlit portrait from a flat midday snapshot, which is the exact distinction the product needs. It then sequences a plan in which each phase removes one of those debts while shipping something visible.
Why the code is private
The reference corpus was scraped from Instagram, which I flagged in the strategy document as a terms-of-service and copyright problem before any commercial use. The architecture treats imagery as a pluggable source so licensed or user-contributed images can replace it. Until that swap happens, the write-up is the artifact and the repository stays private.