Quora Ads ML POC (Light) — Retrieval + Ranking + Brand Safety (Synthetic)

This is a small, readable end-to-end POC aligned to the Quora Ads ML JD:

• Candidate retrieval: two-tower embedding model (in-batch negatives) + Recall@K
• Ranking: CTR model on retrieved candidates (AUC / logloss)
• Brand safety: text classifier + simple safety gating
• Serve demo: serve_request() returns top ads with pCTR + safety score

Run

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
jupyter notebook quora_ads_ml_end_to_end.ipynb