In electronics wholesale, a large share of supply offers is published in Telegram channels and supplier chats: short posts with a model, a price, a quantity and a location. For a buyer following thirty sources, that means hundreds of messages a day and a real risk of missing the best offer.
The pipeline
- Collection. Messages and attached price lists are gathered from the sources the company is a member of.
- Extraction. A language model, guided by examples from your own data, extracts structured fields: brand, model, variant, condition, price, currency, quantity, location, supplier.
- Matching. Extracted items are matched to EAN/MPN codes and to your SKUs, so “iPh 15 128 blk” and a formal part number end up on the same line.
- Ranking and alerts. Duplicates are removed, offers ranked, and buyers notified when targets are met.
What makes it work
The difficult part is not the AI call — it is the quality loop. We measure extraction accuracy on a labelled sample, flag uncertain records for human review and feed corrections back into the rules. Over a few weeks accuracy on routine posts becomes very high, while unusual ones are still checked by a person.
A good system does not replace the buyer’s judgement; it removes the hours spent scrolling.
Staying on the right side of the rules
Only sources the company legitimately has access to are processed, platform terms are respected, and personal data in messages is minimised and protected under GDPR.
Need help with this? Talk to our team.
This article is general information, not legal or tax advice.