Qalxan: scam campaign radar for mobile operators
The user is the anti-fraud analyst at an Azerbaijani mobile operator (Azercell, Bakcell, Nar). Scam SMS reach thousands of subscribers before anyone reacts: fake Kapital Bank "your card is blocked" links, the fake Azerpost 10.2 AZN customs fee, fake Tap.az buyers, "Mom, this is my new number". Today the defence is a keyword blocklist, updated by hand after complaints, with hours to days of lag. Scammers beat it by swapping one letter (ə→e, a Cyrillic "а" inside "Kapitаl"), rotating domains and rephrasing. In our tests the keyword approach catches only 30% of scams. Subscribers lose money, and the operator takes the complaints, the brand damage and the regulator pressure. Qalxan reads every SMS like an analyst. It catches the scheme whatever the wording, groups messages into a campaign within seconds (brand impersonated, domains, phones, messages per minute), and lets the analyst block the campaign and warn everyone who already received it in one click. Every verdict is explained in Azerbaijani, Russian and English. Subscribers can forward a suspicious SMS to a Telegram bot and get the same explanation.
Preliminary score
- Value for the user22.5 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI25.5 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing17.5 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility12.0 / 15
Data requirements, running costs and a clear next step.
- Originality7.5 / 10
A meaningful difference from common solutions.
A finalist: the live jury scored it 71.0 / 100 on stage.
What the judges wrote
- GPT-6 AstraPreliminary84/100
- Strengths
- Your operator-focused workflow connects multilingual SMS detection to campaign triage and block/warn controls rather than stopping at a verdict. Your local classifier plus selective Gemini escalation gives AI a concrete role while limiting API use. Your two test sets, stronger ML baseline, specific failures and explicit disclosure of optimistic retesting provide substantial evidence of thoughtful engineering.
- To improve
- Your next validation should use fresh, consented real SMS and an operator's actual filter, with per-language confusion matrices and uncertainty around false-positive rates. Show completed block/warn effects, distinguish simulated actions from firewall integration, and test incorrect campaign merging. Quantify costs per million SMS and tail latency, including your proposed self-hosted second-opinion model. Define data retention and a review path for risky messages when Gemini fails, rather than simply retaining a potentially wrong local verdict.
- Claude Opus 5.5Preliminary86/100
- Strengths
- Your user (the anti-fraud analyst at Azercell, Bakcell or Nar) and the pain are concrete: hand-updated keyword blocklists that lag by hours and are beaten by letter swaps. The live dashboard shows the core scenario end to end, from traffic stream to campaign card to block and warn. Your own TF-IDF and logistic-regression model handles 97.6% of traffic, with Gemini as a targeted second opinion, which is a thoughtful and cheap architecture. Your testing is unusually honest: two test sets, a keyword baseline and a classic n-gram baseline, documented failures (AnyDesk and card-data requests, the Netflix false alarm), and an explicit caveat that the post-fix 97.5% was measured on the same set that revealed the failures.
- To improve
- Both your training data and Test 2 are generated from your own templates, so the 96–98% figures likely overstate real-world performance; even a few dozen real forwarded SMS would make the evidence much stronger. Your keyword-filter baseline is your own reconstruction, so validate it against an operator's actual blocklist rules during the pilot. On deck slide 7 the bars are labelled only 1.0/2.0/3.0, so name each system on the chart. Also give a rough license price range or cost-per-million-SMS estimate so operators can judge ROI.
The team
Mir Said Hasanzade
product
Emin Bayramov
developer
Farid ValiyevLead
developer