Ctrl + AI
User 1: the mobile subscriber A subscriber of Azercell, Bakcell or Nar calls with a routine question: What is my balance? How much internet do I have left? Can you add a package? Why is my internet slow in my area? Today, this person goes through a menu and often waits in the same queue as people with serious problems, just to hear one number. Two more problems go unnoticed: • Overpaying. Our demo subscriber pays 30 manat for 100 GB and uses 8 GB. Nobody tells him. • Repeating yourself. After a network complaint, the subscriber gets no reference code, so the problem must be explained again on every call. User 2: the call center manager Managers have no quick way to see why people are calling, which districts complain the most, or how long complaints stay open. Why this has not been solved Global voice assistants such as Siri and Alexa do not support Azerbaijani, so an operator cannot simply buy one and switch it on. A solution has to be built for Azerbaijani speech and for each operator's own data. What we built A voice assistant that answers routine questions in seconds, completes package and tariff changes within the same call, tells the subscriber when a cheaper tariff fits their real usage, gives every complaint a reference code, and shows managers live call topics, complaint districts and open-complaint times.
Preliminary score
- Value for the user20.5 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI24.0 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing13.5 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility8.5 / 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 5.5Preliminary72/100
- Strengths
- Your Azerbaijani voice agent goes beyond FAQ answers: it performs real account operations such as package activation, tariff change and complaint codes against a live PostgreSQL backend, and the 30-manat/100 GB overpaying example makes the user value concrete. Your three-path router is a strong engineering choice: deterministic code for account data (~0.1 s), hybrid BGE-M3/FAISS/BM25 for FAQ, and the LLM with tools only for open questions. Code that blocks a 'done' message unless the database write succeeded addresses a real LLM failure mode. Your testing notes are honest and specific, including the Gemini timeouts, Whisper mishearing Azerbaijani, and small talk matching an FAQ at 0.73 against a 0.70 threshold.
- To improve
- Turn your scripted dialogues into a measured test set. For example, run 50–100 Azerbaijani utterances and report intent-routing accuracy, STT error rate, task completion rate, and how many calls still needed a human. Compare against today's IVR flow with a number, such as current menu time or queue wait for a balance question versus your 0.1–4 s. Add a cost per call (STT, LLM and TTS calls) and a concrete next step, such as which operator data or API integration you would need for a pilot. Your submission's setup instructions still contain template placeholders; the README covers this, but filling them in would help reviewers.
- GPT-6 AstraPreliminary75/100
- Strengths
- Your Azerbaijani voice workflow addresses concrete subscriber needs and goes beyond FAQ answers to account changes, cheaper-tariff recommendations and traceable complaints. Your documented separation of deterministic account operations from LLM responses, including database-write verification before reporting success, is a strong design choice. Your testing notes identify specific failures—misrouted small talk, speech-recognition errors and false completion messages—and explain the corresponding fixes.
- To improve
- Your evaluation should report dialogue counts, completion rates and full speech-to-speech latency against the current IVR or agent workflow; the 0.1-second account-processing time is not the total spoken response time. Add documented regression results for noisy or accented Azerbaijani, failed purchases and duplicate transaction requests. Before an operator pilot, specify caller authentication, customer-data handling and billing-system integration, and estimate cost per completed call at realistic volume.
The team
Murad AsadzadeLead
product
Elshan Shukurov
developer
Orkhan Nuriyev
data