LuneX
Primary user: customers of electronics retailers, from small Instagram/WhatsApp shops to large national chains. They want to buy or trade in any technical device: phones, laptops, tablets, smartwatches, headphones, consoles. They reach the store by phone call or WhatsApp. Business user: electronics retailers of every size. Small and mid shops have no dedicated support team. Large chains run expensive call centers with queues. What goes wrong today Example: Aysel wants to swap her old laptop for a new one. She calls the store and hits a phone menu (“press 1, press 2”), then a queue. When she reaches someone, they can’t price her old device over the phone: “Bring it to the store.” She writes on WhatsApp instead and waits hours for a reply. The next day she calls again, and nobody remembers her. For the customer: Waiting in menus and queues, often while paying for the call No trade-in price without visiting the store Repeating the same details on every channel For the retailer: Lost sales: customers buy wherever answers first Small shops: the owner answers messages by hand and misses orders Large chains: high call-center costs, long queues at peak times Trade-in risk: prices are based on what the customer says (“battery is 95%”, “no scratches”), which is often wrong Phone and WhatsApp are disconnected, so there’s no shared customer history Why existing tools don’t solve it Phone menus and basic chatbots route the customer; they don’t complete the sale. General AI chatbots can make up prices or be talked into discounts, and they don’t check device photos.
Preliminary score
- Value for the user19.5 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI24.5 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing16.5 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility11.0 / 15
Data requirements, running costs and a clear next step.
- Originality7.0 / 10
A meaningful difference from common solutions.
A finalist: the live jury scored it 67.0 / 100 on stage.
What the judges wrote
- Claude Opus 5.5Preliminary80/100
- Strengths
- Your core loop is concrete and complete: one agent handles voice and WhatsApp with shared memory, 17 tools, photo-verified trade-in pricing and payment confirmation. Money decisions are enforced in backend code rather than by the model, which directly answers the 'LLM talked into a discount' risk. Your testing is unusually honest and systematic: 30 simulated-customer scenarios scored from DB and tool logs, 143 backend tests including a 1,000-case fuzz of the 5% cap, and a candid audit of a 20% first run that traced the failures to your own evaluator. The cost model (per-minute voice, per-chat, per-photo, Twilio rates) against a 2–3 operator baseline is specific enough to check.
- To improve
- Report real-photo vision accuracy with numbers: how many of your own device photos had battery or damage read correctly, and where it failed, since agent evals used mocked vision. Finish more of the 54 planned eval runs and test the Twilio phone line live end to end, because the real telephone path is what retailers will actually adopt. Add a business metric from a real shop, even an informal one such as current response time or missed WhatsApp orders per day, so the outcome is measured rather than asserted. On the deck, Solution 03 duplicates Solution 02; use that space for the human-handoff or memory feature.
- GPT-6 AstraPreliminary77/100
- Strengths
- Your electronics-retail workflow addresses concrete friction by connecting inventory lookup, photo-assisted trade-in, negotiation and sandbox checkout across voice and chat. Your separation of AI conversation from code-enforced pricing and payment state is a strong engineering choice, and you clearly distinguish the scripted demo from the live agent. Your reported backend tests, log-based evaluations and negotiation-cap fuzzing target important financial failures, while the disclosed search bugs and incomplete coverage make the testing account more credible.
- To improve
- Your next validation should document a reproducible end-to-end run through real telephone and WhatsApp channels, clearly labelling synthetic inventory and simulated payments. Publish a reconciled results table explaining the 90 audited conversations, 19/54 completed runs and 16/16 passing subset, alongside real-photo accuracy tests against human inspections. In your planned shop pilots, compare quote accuracy, sale completion and staff time against the existing workflow. Reconcile your monthly cost estimate: at the stated rates, 9,000 voice minutes and 5,000 chats already cost $1,300 before telecom, callbacks, hosting and human handoffs.
The team
Atilla Mustafa
developer
Omar Sultanli
developer
Ali Safarov
business
Elkhan Aghakishiyev
data
Inji AbdulazizovaLead
product