On a marketplace, trust is the product. Theirs was being manufactured: bots and paid actors flooded the site with fake reviews, planted ratings and content that should never have reached a shopper. The old filters missed it, and a human team could not keep pace.
A marketplace runs on the reviews and ratings customers leave. Theirs were under attack: coordinated actors and bots posted fake five-star reviews to prop up bad sellers, and fake one-star reviews to sink honest competitors. The signal buyers relied on was quietly turning to noise.
Harmful content made it worse. Offensive or infringing images and text were uploaded against products and shown to the public before the basic filters could catch them, putting the brand at risk every time. And the team could not keep up: the trust-and-safety group was pouring its budget into manually reviewing millions of transactions and comments, and still falling behind the growth of the site.
The old automated checks were no match for sophisticated manipulation. They caught the clumsy attempts and missed the coordinated ones, which are exactly the ones that do the damage.
An engine that reads the patterns humans cannot, with people on the calls that matter.
We wired in an engine that weighs thousands of signals at once, the network and device fingerprint, the wording, the pace of posting, the reviewer's history, and flags a fake review or abnormal item in a fraction of a second.
High-confidence fakes and harmful content are filtered and removed automatically, roughly 90% of the volume, before a shopper ever sees them. The obvious abuse no longer waits in a human queue.
Complex items and suspicious sellers are routed to a trained review team for a final call, so a borderline decision is made by a human, not guessed by a model under pressure.
Every human ruling is captured into a clean, verified dataset, and we use it to retrain the engine so it recognises new fraud patterns on its own the next time they appear.
Offensive and infringing content is screened before publication rather than pulled down after complaints, so the harm is prevented instead of cleaned up.
Because the engine absorbs the routine load, moderation stops scaling one-to-one with the size of the site, and the human team is focused where judgement actually matters.
Fakes stopped at the door, and the cost of stopping them fell.
More than 95% of fake reviews and inappropriate content was removed or blocked before it reached customers, which protected the platform's reputation and lifted the trust scores buyers rely on, along with purchases from properly rated, trustworthy sellers.
Automating the filtering of content and the vetting of seller ratings produced a major reduction in monitoring, review and support costs, and let the trust-and-safety team keep pace with a growing site without growing one-to-one alongside it.
Moderating after the fact means the damage is already done. The value here is catching abuse before publication, drawing a hard line at high confidence, and sending only the genuine doubt to a person, whose decision then teaches the model. It is the same grounded, human-on-the-hard-cases discipline behind our fraud-detection work, aimed at content and reputation rather than transactions.
Toxic chat was driving players off a gaming platform while regulators circled. Context-aware moderation across 30+ languages told banter from abuse, cutting toxicity 45%, lifting new-player retention 20%, and dropping false blocks 75%.
Read the case → Credit scoring · AgricultureApproving farm credit meant sending a surveyor out for weeks. Computer vision on satellite imagery scored crop health and collateral, taking the company from dozens of applications a week to 25,000 farms a day and cutting underwriting cost over 70%.
Read the case →
Reef TRH