MISSION REPORT · Build · 2025-03
Multimodal Image-Text Retrieval (CLIP fine-tune)
Fine-tuned CLIP on 5–10k pairs with contrastive loss; hard-negative mining; FAISS ANN search.
+14–20%Recall@10
< 100 msCPU query latency
- PyTorch
- CLIP
- FAISS
DOSSIER
Mission report in progress
This mission report is still being written. The archive one-liner above is the summary on file.
Fixed-query retrieval
Illustrative demo with an off-the-shelf CLIP model — not Distill-CLIP.
A fixed list of phrases is matched to a committed set of public-domain thumbnails by cosine similarity. Nothing is downloaded into the browser except those files.
