AICell+
Who: prepaid Azercell subscribers who manage their line in the self-care app on their phone, often one-handed or on the move. What goes wrong today: - The common jobs are checking the balance, checking data and minutes left, topping up, buying an internet pack and changing tariff. The app spreads these across five tabs and dozens of screens, so every job means knowing which screen to open. - Nothing in the app tells subscribers when they pay too much. In our seeded data, one subscriber who keeps buying packs pays 5.98 AZN more than the same data would cost on the next tariff up. - Support questions mean searching the website or waiting for an operator. Outcome: the subscriber says "Hey AIsel" and the request in Azerbaijani or English. The app does the job through the same screens a person would tap, shows each step with a visible cursor, and speaks the result. It also says unprompted when a tariff change would save money, and a spoken "yes" makes the change. Payments still need the phone's fingerprint (or FaceID). Link to presentation just in case: https://docs.google.com/presentation/d/1i0SqIFMjgEn9YVkuBpLcRJ53x5DHjAwK/edit?usp=sharing&ouid=102509767929492463555&rtpof=true&sd=true
Preliminary score
- Value for the user19.0 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI22.5 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing15.5 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility7.5 / 15
Data requirements, running costs and a clear next step.
- Originality7.0 / 10
A meaningful difference from common solutions.
What the judges wrote
- Claude Opus 5.5Preliminary70/100
- Strengths
- Your core path is concrete and well scoped: the assistant reads per-day and per-app usage, prices the overpayment in manats, and a spoken 'yes' makes the change, with biometric confirmation kept for payments. The 'numbers from code, words from AI' guard is a smart design choice, and you back it with tested safety rules: invented numbers are replaced, out-of-list tasks are rejected, and a model failure falls back to a fixed sentence. Your router comparison (Convai Laya 8/17 vs Gemini 17/17 on Azerbaijani) and the specific failures you report, including the TTS style instruction read aloud and '$21.21' from speech to text, are exactly the honest, evidence-based testing judges want. You also say plainly that you did not time voice against tapping, which makes your other claims more credible.
- To improve
- Your track asks for a metric that justifies adoption, and you have not made the business case for Azercell. Explain why an operator would deploy a tool that steers subscribers towards cheaper options, for example through churn reduction, fewer support calls, or NPS, and estimate that effect. Feasibility needs running costs: give per-conversation Gemini STT, TTS and LLM cost, plus a realistic pilot as your next step. Run the voice-versus-tap timing test on even 5–10 people, and grow the 31-message router set with real-sounding Azerbaijani phrasing. Also reconcile your test counts: the deck says 1,495 tests pass and the submission says 1,443.
- GPT-6 AstraPreliminary73/100
- Strengths
- Your project addresses concrete prepaid-subscriber problems, with a measurable savings example of 5.98 AZN in the seeded data. Your separation of deterministic financial calculations from AI-generated language, together with allowlisted actions and confirmation safeguards, is a sensible design. Your testing report provides specific bilingual routing failures, speech issues, and corrective changes rather than relying only on a large passing-test count.
- To improve
- Your next demonstration should document a complete voice-to-confirmation-to-balance-update trace in both languages, clearly distinguishing simulated transactions from live integrations. Expand the 31-message router evaluation into held-out, end-to-end voice tests, including currency ambiguity, duplicate confirmations, and transaction failures. Measure task completion, time, or support deflection against the current app, and validate projected savings beyond seeded examples. Add per-interaction running costs and a realistic operator pilot plan covering consented usage access, production authentication, and payment integration.
The team
Shahnaz Orujlu
developer
Riad Mukhtarov
developer
Gudrat AbidzadaLead
developer