Awas
An anti-scam advertisement checker that extracts advertisement fields, matches fixed rules, and cites official pages to generate an evidence file. It outputs one of three conclusions—STOP, VERIFY FIRST, or NO RED FLAGS FOUND—rather than a risk score.
Timeline (last 90 days)
Developer edycutjong built the Awas project for the Nebius global AI hackathon.
Awas's user flow is: the user pastes ad text, and Awas returns an evidence file and one of three verdicts (STOP, VERIFY FIRST, or NO RED FLAGS FOUND), with no score.
Awas uses Nemotron 3 Ultra on Nebius Token Factory to extract structured fields from ads and uses Nemotron 3 Super to propose evidence rows.
Awas's citation gate is implemented in pure code, keeping only those evidence rows whose citations are verbatim substrings of page text fetched during that run.
The coverage gate set in advance by the author requires that, for at least 12 of 20 known scam cases, Tavily must find a relevant official warning or news item.
The coverage gate benchmark run found only 8 of the 20 cases, failing to meet the pass condition.
Among the benchmark's reported cases, 1 case in the dev set, 0 in the first batch, and 1 in the second batch mentioned the company name.
After the coverage gate failed, the author redefined the product scope: verdicts come from rules citing official pages, and web search serves only as background information for family members, never as an input to the rules.
In the holdout run, the number of false stops by Awas on job openings at 10 licensed institutions was 0.
In the holdout run, Awas gave 9 STOP, 15 VERIFY FIRST, and 1 NO RED FLAGS FOUND on 25 publicly reported recruitment scam cases; eight of the nine STOPs came solely from cited rules.
3 more timeline events
All Facts (13)
Developer edycutjong built the Awas project for the Nebius global AI hackathon.
50%UnverifiedAwas's user flow is: the user pastes ad text, and Awas returns an evidence file and one of three verdicts (STOP, VERIFY FIRST, or NO RED FLAGS FOUND), with no score.
50%UnverifiedAwas uses Nemotron 3 Ultra on Nebius Token Factory to extract structured fields from ads and uses Nemotron 3 Super to propose evidence rows.
50%UnverifiedAwas's citation gate is implemented in pure code, keeping only those evidence rows whose citations are verbatim substrings of page text fetched during that run.
50%UnverifiedThe coverage gate set in advance by the author requires that, for at least 12 of 20 known scam cases, Tavily must find a relevant official warning or news item.
50%UnverifiedThe coverage gate benchmark run found only 8 of the 20 cases, failing to meet the pass condition.
50%UnverifiedAmong the benchmark's reported cases, 1 case in the dev set, 0 in the first batch, and 1 in the second batch mentioned the company name.
50%UnverifiedAfter the coverage gate failed, the author redefined the product scope: verdicts come from rules citing official pages, and web search serves only as background information for family members, never as an input to the rules.
50%UnverifiedIn the holdout run, the number of false stops by Awas on job openings at 10 licensed institutions was 0.
50%UnverifiedIn the holdout run, Awas gave 9 STOP, 15 VERIFY FIRST, and 1 NO RED FLAGS FOUND on 25 publicly reported recruitment scam cases; eight of the nine STOPs came solely from cited rules.
50%UnverifiedIn the holdout run, all 48 displayed evidence rows passed the citation gate; 6 of 54 proposed rows were dropped (11%).
50%UnverifiedThe server-side performance of the Awas holdout run was p50 13.7 seconds and p95 31.7 seconds; the median cost was $0.0078 of Token Factory tokens plus 1 to 3 Tavily credits.
50%UnverifiedIn the holdout run, across 20 reruns on 10 already-reported cases, the verdict was identical in 20 runs; across the 10 licensed job openings, the verdict was identical in 19 runs.
50%