cortexa-marketing-scorer (v10)

Production scorer for Pleius Marketing. Given an ad image and a target demographic, returns 4 z-scored pillars:

• universal_appeal — does it work for everyone
• demographic_appeal — does it work for the target slice (amplified against a "general audience" baseline; see inference.py in the serving Space M725/cortexa-ad-api)
• audience_drive — likelihood that the target acts on the ad
• engagement — saveable / shareable / rewatchable

Pipeline

ad image ─→ CLIP ViT-B/32 vision encoder (fp16 ONNX) ─→ (1, 512) stim_emb
demographic ─→ CLIP text encoder       (fp16 ONNX) ─→ (1, 512) demo_emb
                                  ↓ concat
                              cortexa_v10 head (ONNX)
                                  ↓
                         4 metric heads → z-scores

Demographic amplification (×12 over the "general audience" baseline) is documented inline in the serving Space's inference.py.

Files

file	purpose
clip_vision_fp16.onnx	CLIP ViT-B/32 vision tower, fp16. Disable graph opts (SimplifiedLayerNormFusion bug with fp16 Cast graphs).
clip_text_fp16.onnx	CLIP text tower, fp16. Same caveat.
cortexa_v10.onnx	Pleius-trained metric head. Takes (demo_emb, stim_emb) → 4 metric tensors in METRICS order.
clip_tokenizer.json	HuggingFace tokenizers JSON. Pad to 77, truncate at 77.

Reference inference

See M725/cortexa-ad-api's app/inference.py::CortexaPipeline for the canonical CPU inference path.

License

Pleius internal — see https://pleius.com. Not for redistribution.