Axon
The answer exists. The agent can’t find it in time. The user: a call center agent who handles dozens of calls a day in Azerbaijani and Russian and must know hundreds of rules on returns, refunds, credit and delivery. Searching while the customer waits. Agents put customers on hold to search the FAQ or ask a colleague. Wrong answers come back. About 1 in 6 complaint calls is a repeat call about the same issue. Problems surface late. Only a sample of calls is reviewed, so managers learn about new issues after many customers are affected. Agents don’t lack knowledge. They lack time to find it. Axon finds the answer while the customer is still talking. The solution: an AI copilot that listens to the call in Azerbaijani and Russian, understands what the customer needs and shows the right answer from the company’s FAQ on the agent’s screen. No more searching. As the customer speaks, Axon shows a suggested reply and the next step in 1–2 seconds, with the FAQ rule it comes from. The right answer the first time. Every hint follows the company’s own rules on returns, refunds, credit and delivery, so new and experienced agents give the same correct answer. Every call counts. After each call, Axon labels the topic, the outcome and the customer’s mood, so managers see new issues across all calls, not a sample. Axon suggests. The agent decides.
Preliminary score
- Value for the user20.0 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI19.5 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing11.5 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility12.5 / 15
Data requirements, running costs and a clear next step.
- Originality6.0 / 10
A meaningful difference from common solutions.
What the judges wrote
- GPT-6 AstraPreliminary67/100
- Strengths
- Your focus on Azerbaijani–Russian call-center agents identifies a specific user and concrete pain, while FAQ-linked hints keep agents in control. Your README clearly documents the post-call and replay workflows, including live/cache/fallback indicators and timeout handling. Your itemized running-cost assumptions and four-week pilot plan provide a concrete path toward business validation.
- To improve
- Your next demonstration should process an unseen bilingual call with uncached model outputs and measured latency, clearly distinguishing replay from genuine streaming. Reconcile the deck’s real-call benchmark with the repository’s all-synthetic data declaration, and publish test-set sizes, model outputs and FAQ-answer accuracy separately from mocked software tests. Measure answer correctness and time-to-answer against manual FAQ lookup rather than relying on productivity gains from another study. Validate the optimized API-cost assumptions and explain how sending audio to Gemini fits your local-data-hosting and privacy promises.
- Claude Opus 5.5Preliminary72/100
- Strengths
- Your user and pain are sharply defined: an Azerbaijani/Russian call center agent who must find FAQ rules while the customer waits, with a clear 'Axon suggests, the agent decides' principle. Your README describes a real FastAPI + Gemini backend with transcribe/analyze/recommend endpoints, FAQ-grounded hints that cite their source, stereo channel splitting, JSON repair, timeouts with cache fallback, and 17 tests. The 8-model WER comparison with cut-off, bad-JSON and looping counts is concrete evidence for your model choice. Your feasibility work is unusually thorough: per-minute cost breakdowns, cost-reduction levers, break-even, and a measurable 4-week pilot.
- To improve
- Your materials disagree on the test data: slide 9 says 'real calls with hand-made reference transcripts', while the README and submission say all audio is synthetic text-to-speech. State exactly what the WER set was and how many calls it covered. Measure the core value directly: FAQ-match accuracy on a labelled set of customer questions, and the real hint latency (the deck says under 1 s, the submission says 1–2 s, the README target is under 4 s). Template speaker notes such as 'replace the examples with what actually broke' make it unclear whether the failure table is real, so attach logs or examples. The live copilot is a replay of a recorded file, so show a true streaming run or name it as the next milestone.
The team
Elshad Isgandarli
developer
Murad Mammadzada
data
Nazrin MalikovaLead
product