Problem
An ad-spend analytics company must label every ad landing URL with a brand, an advertiser and a category, because client reports cannot be built without them. The URL alone does not reveal them, so the page must be parsed first. Manual labeling works today, at 37 to 500 new base URLs a day per product. A planned panel of 1000 users would bring about 50 thousand target links per day, roughly 20 days of manual work. The answer must come within 3 seconds on CPU, since GPUs are too expensive for this task.
Illustrative example: a banner links to an online shop. The service returns its brand, the legal entity behind it (the advertiser) and a category.
Solution in brief
A browser-extension parser collects page metadata, text and the privacy policy. A fine-tuned DeepPavlov classifier and brand NER run on CPU. A cheap LLM (DeepSeek-chat) is called only when NER confidence is below 85% and for the advertiser name.
Questions this case answers
- How to label items with entities extracted from their linked pages using small CPU models, calling an LLM only when the model is unsure, within a strict per-request latency limit
- How to collect the page features a classifier needs when server-side collection gets blocked
- How to attribute ad landing URLs to brand, advertiser and category automatically at scale
- How to add an LLM fallback to a small CPU-only NER and classifier pipeline using a softmax confidence threshold
Tools and methods
- Browser extension (content script) + RabbitMQ with Web STOMP — parsing on the page itself, results through queues
- DeepPavlov classifier and NER — category and brand; multilingual BERT as the comparison
- Softmax confidence threshold of 85% — decides when to call the LLM
- DeepSeek-chat via API — fallback for the brand and the main method for the advertiser
- Dynamic quantization, Tesseract, Hugging Face Hub — CPU speedup, PDF policies, model delivery
Google or a one-time email code. No password.
