Rabia Yoruk

Scrapus

Web monitoring for the Turkish Pharmacists' Association (TEB). It finds banned medicine listings on online marketplaces and sends them to TEB's systems.

Role
Co-founder
Where
Scrapus, Tallinn
When
Feb 2025 to now
Stack
  • Python
  • FastAPI
  • LLM agents
  • RAG
  • REST API

60,000

records processed per day for TEB, on average

90%

precision and recall for lead qualification, approximately, as reported in the Frontiers paper

0.77 to 0.92

extraction F1 with transformer NLP, reported in the paper

Scrapus pipeline for TEB, simplified. Flow: Marketplaces and Classified ads, then Crawlers, then Normalize, then LLM classify and Reference corpus, then REST API, then TEB systems. Marketplaces public listings Classified ads public listings Crawlers scheduled runs Normalize clean, dedupe LLM classify enrich, qualify Reference corpus retrieval (RAG) REST API flagged records TEB systems client Scrapus pipeline for TEB, simplified. Flow: Marketplaces and Classified ads, then Crawlers, then Normalize, then LLM classify and Reference corpus, then REST API, then TEB systems. Marketplaces public listings Classified ads public listings Crawlers scheduled runs Normalize clean, dedupe LLM classify enrich, qualify Reference corpus retrieval (RAG) REST API flagged records TEB systems client
Fig. 1 Scrapus pipeline for TEB, simplified.

Problem

Banned medicines are sold on online marketplaces every day. New listings appear faster than people can check them. TEB needed a system that finds them and sends them into its own tools, without a person watching the pipeline.

What I built

  • The whole pipeline, end to end: collection, cleaning, LLM-based classification and enrichment, and API delivery.
  • LLM agents with retrieval (RAG) over a domain corpus. The model sees the listing and the relevant reference entries before it decides.
  • Delivery through an API, so flagged records arrive in the client’s systems and not in a spreadsheet.

The pipeline runs at about 60,000 records a day with minimal maintenance.

From one client to a product

The same engine now powers a general B2B data product with four packages: FindCustomer, FindService, FindJob and FindTalent. Users pay per record collected. A run that finds nothing costs nothing. The approach is described in a paper in Frontiers in Artificial Intelligence (2025). The FindJob agent is described in a separate D3AI paper.

What I would measure next

Cost per thousand records, accuracy against a labelled sample, and drift over time. This is the topic of a paper in preparation.

Scrapus app, Packages page. Four packages: FindCustomer, FindService, FindJob and FindTalent, each with a credit cost per record.
Fig. 2 The general product. Each package is a scraping job with a cost per record. Demo account.
Scrapus app, Learn page with tutorials on reading result rows, writing prompts and saving credits.
Fig. 3 In-app guides. Users learn to read a result row and write better prompts.