An AI agent that watches competitors' online stores: on-demand scans with LLM analysis and a live progress panel, plus recurring product intelligence that spots market-wide promotion trends.
Built for a retail e-commerce company in Poland
The marketing team needed to know what competitors were promoting — but checking each store by hand was slow, irregular, and the first thing to be skipped in a busy week.
Now an employee picks competitors in a browser panel and starts a scan. The agent fetches each store's pages, has a language model analyze them into four fixed sections — campaigns, promotions, new products, other observations — and streams progress and results live into the panel. One click exports everything as a shareable PDF report.
A second track runs on a schedule: it probes competitors' promotion pages, extracts which products are being pushed and how, and flags products and topics — think vitamin D, collagen, magnesium — promoted by several competitors at once. That's the difference between one shop's clearance and a market-wide trend.
Two use cases on one scraping foundation: an on-demand competitor scan with live progress, and a recurring product-intelligence pipeline that files its evidence step by step.
On-demand scan
Marketing panel
team picks competitors, starts scan
Scraping layer
HTTP + headless-browser fallback
LLM analysis
4 fixed sections + market summary
Live view + PDF
results in panel, exportable report
Product intelligence
Competitor stores
public pages & promo subpages
Source probing
promo keywords & product links
Product extraction
LLM + rules, normalized names
Common products
topics shared across competitors
PostgreSQL
task, live log, results · gray = shared scraping foundation
Started by the team, whenever needed
Recurring, in separately runnable steps
Competitor sites change, block bots and go down — the agent is designed around that, so the team gets an answer anyway.
The panel is visible only to the marketing group and administrators.
Automatic retries on server errors, and a headless-browser fallback when a site blocks plain HTTP requests.
Every scan has its own record with progress and log, live in the panel — regardless of how many ran before.
The LLM must answer in four fixed sections, so a clean PDF can always be built from the result.
The intelligence pipeline is split into probe → extract → compare; each step saves its output and can be re-run alone.
Every observation traces back to raw source files — you can always see where a conclusion came from.
Technology: Python · Django · LLM APIs · headless Chrome/Selenium · PostgreSQL
Competitor monitoring, price watching, trend detection — an agent can check every day what a person checks once a quarter. Let's talk.
Discuss your case