DuyAI
WHO: people who must handle a real phone call in a language they don't speak — students and migrants abroad, older people calling banks and clinics, hotels and small businesses taking calls from foreign clients. PROBLEM TODAY: translator apps work face to face, not inside a phone call. Speakerphone + translator app is slow and the other side hangs up; a human interpreter is expensive and must be booked. After the call nobody remembers exactly what was agreed (date, time, price, who promised what). Phone scammers target people who don't fully understand the language. OUTCOME WITH DUYAI: you call from your own SIM and number, the other person installs nothing. AI translates both directions by voice in 21 languages, shows live captions, warns about scams during the call (before you read out an SMS code or card number), and turns the call into a report with calendar events, reminders and "who promised what". PROOF: ~600 people tried it (hackathon visitors, friends and family), 94% found it useful. Dates, times and prices extracted with 100% accuracy on our hand-labelled Azerbaijani test set. WHY US: the only solution that works inside real calls on ANY phone, handles Azerbaijani mixed with Russian/English/Turkish, turns calls into actions and guards against scams. We also trained our own small AI model for Azerbaijani calls (offline/on-device path). Accepted into Anthropic's Claude for Startups programme during the hackathon; won a 1-year Replit Core plan for the team.
Preliminary score
- Value for the user19.5 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI23.5 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing15.0 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility11.0 / 15
Data requirements, running costs and a clear next step.
- Originality6.5 / 10
A meaningful difference from common solutions.
What the judges wrote
- Claude Opus 5.5Preliminary77/100
- Strengths
- Your prototype tackles a hard, concrete problem: two-way voice translation inside a real GSM call where the other party installs nothing. You describe a working Azure Speech/Translator/TTS pipeline with Claude for context-aware reports, commitments and scam risk, plus a public demo and video. Your testing section stands out. It includes a hand-labelled Azerbaijani test set with per-field metrics and 84 automated tests. It also lists real call failures you hit and fixed: Bluetooth audio capture, echo, Azerbaijani/Turkish confusion and Wi-Fi drops. You also compare latency and cost against speakerphone-plus-translator and human interpreters, about 2.5–4.5 s per phrase and $0.07 per minute.
- To improve
- Your evaluation set is only 24 calls that your own team wrote, and the scam check covers just 10 phrases. 100% scores on that data say little about messy real speech, so test on recorded real calls with real ASR errors and report end-to-end translation quality. The claim that ~600 people tried it in one day with 94% finding it useful needs a method: what was asked, how it was recorded, and how many completed full calls. The full setup currently depends on a Windows PC with Phone Link, VB-CABLE, Voicemeeter and Stereo Mix, which is a big adoption barrier. Spell out a realistic path to a phone-native or telephony-based (e.g. Twilio) version, and address call-recording consent. Your market slide targets import/export agencies while your user story centres on migrants and older people, so pick one primary customer for the enterprise track.
- GPT-6 AstraPreliminary74/100
- Strengths
- Your core scenario—an Azerbaijani speaker calling a Russian-speaking landlord, clinic or hotel—connects a real communication barrier to concrete calendar and follow-up actions. Your documented Phone Link pipeline gives AI a substantive role in live translation and post-call understanding, without requiring recipient installation. Your 24-call evaluation, small-model before/after results and specific audio-routing failures provide useful testing detail beyond generic success claims.
- To improve
- Your next validation should focus on one business workflow, measuring task completion, misunderstandings and end-to-end latency against a translator-app baseline. Your 24 written conversations mainly test call understanding; add unseen recordings with noise, accents and code-switching, plus benign scam-like requests, and report missed alerts and false positives. Your 'any phone' positioning should distinguish the current Windows + Android setup from the simulated-call demo and planned on-device model. Your feasibility case would be stronger with a breakdown of the $0.07/min estimate and consent, cloud-processing and deletion safeguards in place before pilots.
The team
Rafael Qurbanov
developer
Elgun ShukurovLead
data
Laçın Qayıbova
business
Gulustan Nabiyeva
designer
İlkin Cəfərli
product