Aura
People in Azerbaijan get scam messages every day ("your Azercell bonus is ready, click", "your card is blocked, send the code"). Scammers write in Azerbaijani with and without special letters (translit), in Russian, or in a mix, so keyword filters miss them, and they also flag real bank notices. Banks and mobile operators only learn about a campaign when customers complain. Two users: (1) anyone who gets a suspicious message and cannot tell if it is a scam: they send the text or a screenshot (website or Telegram bot) and get a verdict in Azerbaijani, English or Russian, with the reasons and what to do; (2) fraud teams at banks and telecoms: they see campaigns that impersonate their brand, get a blocklist of fake domains (CSV) and a signed webhook alert when a new campaign appears, through a partner API.
Preliminary score
- Value for the user21.0 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI24.5 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing16.5 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility11.5 / 15
Data requirements, running costs and a clear next step.
- Originality7.0 / 10
A meaningful difference from common solutions.
A finalist: the live jury scored it 78.0 / 100 on stage.
What the judges wrote
- GPT-6 AstraPreliminary82/100
- Strengths
- Your focus on Azerbaijani, transliterated and mixed-language scams addresses a concrete problem for citizens and bank/telecom fraud teams. Your documented workflow connects explainable verdicts and domain checks to campaign grouping, brand-scoped APIs, CSV blocklists and signed alerts. Your reported baseline comparison, latency/cost measurements and specific failure fixes provide strong engineering evidence, while clearly acknowledging the synthetic benchmark's limitations.
- To improve
- Your next evaluation should use independently labeled real scams and legitimate notices, covering screenshots and Russian/English detection against a stronger baseline than keywords. Your enterprise demonstration should trace a fresh submission through campaign grouping, a received signed webhook and a brand-scoped blocklist, measuring alert latency and duplicate-alert volume. Before enabling automatic blocking, verify your official-domain registry and test legitimate new domains against your guardrails. Your rollout plan should include hosted-backend costs, resilience to exhausted vision quotas and verification that screenshot data is redacted before external processing.
- Claude Opus 5.5Preliminary79/100
- Strengths
- Your submission covers both sides of the problem with working pieces: a citizen-facing checker (web plus Telegram bot, text and screenshots, AZ/EN/RU answers) and a business-facing partner API with brand-scoped keys, a CSV blocklist and HMAC-signed webhooks. AI does real work here: an LLM verdict sits behind PII redaction, RDAP domain-age checks and code-level guardrails, and a second model acts as an attacker. Your testing write-up is unusually honest and specific. It compares you against a keyword baseline (88.9% recall with 50% false alarms), lists seven concrete bugs you found and fixed (the İ normalization split, cached RDAP errors, truncated JSON), and states clearly why your 100% score on a self-written set is not proof.
- To improve
- Your 45 test messages and 71 attack variants are all LLM-generated, and the attacker found zero misses. Your next priority should be a set of real, anonymized scam and legitimate messages, for example via add_real.py with a bank or CERT.az, labelled by hand, so the recall and false-positive numbers mean something. The deck goes further than the evidence: it says the model is 'trained' on local text (it is prompted), it shows dashboard figures like 12.4K messages that look like demo data, and it ticks every column against Truecaller and VirusTotal. Keep the deck aligned with what you measured. For the enterprise track, name one pilot workflow and its success metric, such as time-to-detect a new Azercell impersonation campaign versus complaint-driven detection today, and say how you would move the backend off a laptop.
The team
Ilaha Mardiyeva
data
Chichak Asgarova
product
Ravan IsmayilovLead
developer