TelsyAI | Telecom Sales Assistant
We built this for telecom sales teams in Azerbaijan. Customers can keep paying extra for data even when a different package would cost less. Finding these customers, comparing packages and calling them takes staff time. Companies also pay salaries and spend money hiring and training people to handle these repeated conversations. Our aim is to reduce that cost by letting AI explain suitable offers, while employees approve recommendations and handle more difficult cases.
Preliminary score
- Value for the user19.0 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI23.0 / 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.
- Originality6.0 / 10
A meaningful difference from common solutions.
What the judges wrote
- Claude Opus 5.5Preliminary67/100
- Strengths
- You picked a concrete, well-scoped workflow: finding telecom customers who overpay on data overage, comparing packages, and calling them. You keep an employee approval step before and after the call. Your architecture is careful. Recommendations come from explainable Decimal-based rules, GPT-4.1 mini only selects intents and allowlisted facts through structured outputs, and the backend supplies every spoken reply, which sharply reduces the risk of hallucinated prices. The full loop is deployed: approval, voice or typed call, STT, TTS, saved outcome and dashboard. Your testing write-up is honest and specific: 3 of 20 replies misunderstood, the 'I don't need more data' objection bug, a 0.36-second truncated audio clip, and roughly 15-second voice latency.
- To improve
- You have not shown a business metric yet. Measure the time an employee spends on manual bill comparison and calls today against your flow, and estimate conversion or retention impact, because the track asks for the metric that justifies adoption. Add running costs: per-call token, STT and TTS spend at realistic call volumes, compared with staff cost per conversation. Expand testing beyond 20 self-written replies to unseen or adversarial customer phrasings, and ideally a few real or anonymised usage profiles from an operator. Treat cutting the 15-second voice latency, for example by streaming TTS or pre-generating fixed replies, as a concrete next step.
- GPT-6 AstraPreliminary73/100
- Strengths
- Your materials describe a complete recommendation-to-approval-to-conversation-to-saved-outcome workflow, with AI interpreting replies while the backend controls package facts. Your contact-permission checks, verified offer snapshots, and explicit confirmation state are thoughtful safeguards for telecom sales. Your testing identifies concrete intent and audio failures, documents fixes, and honestly acknowledges unresolved latency.
- To improve
- Your next evaluation should use unseen conversations in your target customers’ languages, rather than only rerunning the same 20 replies. Compare staff-led and assisted workflows on handling time, decision accuracy, and completion rate. Measure and reduce the reported 15-second voice delays, and report API cost per completed conversation. Define a small consented telecom pilot with real catalog and usage data, authenticated employees, and a clear handoff for package activation.
The team
Fidan İsgəndərli
designer
Ləman İsazadə
developer
Rahman KarimliLead
developer