AI Enterprise SolutionsGPT & Claude

whoami

Azerbaijani telecom customers — modeled on Azercell's ~5M subscriber base — routinely lose money to two things they rarely notice until it's too late: unwanted recurring third-party charges they never knowingly signed up for, and mobile plans that no longer match how they actually use data. Industry research (C+R Research) shows people underestimate their real subscription spend by more than 2x ($86 estimated vs. $219 actual), and 42% have been charged for a subscription they'd completely forgotten. Azercell's own AI support bot, AiCell, understands Azerbaijani with 96.6% comprehension accuracy but resolves only 17% of inquiries end-to-end — the other 83% get routed to a human, meaning the customer still has to make the call themselves. Clario targets both halves of this: it notices a suspicious charge before the customer does and actually cancels and refunds it, and it recommends a better-fitting plan based on what the customer says about their own usage — closing the loop AiCell leaves open.

Open the demo Video GitHub

Preliminary score

70.5/ 100Strongaverage of 2 judges

Value for the user19.5 / 25
A specific problem and a clear outcome from solving it.
Prototype and use of AI22.0 / 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.
Feasibility9.5 / 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 AstraPreliminary72/100
Strengths
Your tool-calling workflow connects Azerbaijani conversation to concrete billing actions, with confirmation and disambiguation rather than advice alone. Your reported tests cover repeat-cancel idempotency and a thoughtfully corrected zero-context ambiguity case, while the recommender includes cross-validation and usage-drift checks. Your explicit disclosure of mock bills and synthetic training data makes the prototype's current boundaries clear.
To improve
Your evidence would be stronger with tool traces showing both cancellation and refund state changes, plus a case-by-case table separating completed resolutions from clarification-only outcomes. Your 0.992 macro-F1 measures agreement with synthetic personas, not real plan suitability; validate against real usage and a simple cheapest-matching-plan baseline. Your production plan needs operator API access, authentication, refund authorization, and audit controls—not just a replacement data source. Your $0.008 cost estimate should specify the model pricing, token assumptions, and number of calls per resolved case.
Claude Opus 5.5Preliminary69/100
Strengths
Your prototype goes beyond explaining a charge: Claude calls get_bill and cancel_subscription tools, and asks for confirmation before any irreversible action, which is a real agentic loop. Your test write-up is unusually honest. You admit that your first 'ambiguity' test did not test ambiguity, then you replaced it with a true zero-context case with two active charges. Choosing logistic regression so the recommender runs client-side with no inference cost is a thoughtful engineering trade-off. Your next step of swapping mock data for a real billing system is clearly scoped.
To improve
Your 0.992 macro-F1 comes from predicting K-Means labels on the same synthetic features, so it mostly shows the classifier can recover its own clusters. It does not show that real users get better plans. Validate on even a small set of real usage descriptions or anonymised bills. Your comparison with AiCell's 17% is not like-for-like: measure your own resolution rate, time-to-resolution or AZN recovered against the current call-centre path on the same scenarios. Your problem statistics are US subscription data, so local evidence on unwanted VAS charges in Azerbaijan would strengthen your user-value case. Also address how refund and cancellation authority would work under operator policy.

The team

Aqshin Quliyev
developer
Hamid AslanovLead
developer
Nicat Əmrəhli
product
Elvin Bayramli
data
Vagif Akbarli
developer

Disclosure

Required by the rules. Judges verify it against the repository.

Models: Anthropic Claude Sonnet 5, scikit-learn LR multinomial, scikit-learn, KMeans
Data: Syntetic mock bill data
Components: Next.js, React 19, Typescript, Tailwind CSSv4, Vercel, GitHub
Built during the event: Confirmed

Quality testing

Billing-dispute flow — 11 real test cases, run against the live Claude Sonnet 5 API (not simulated): vague opener, direct cancellation, repeat-cancel idempotency, off-topic small talk, slang/colloquial complaints, the proactive opening trigger, and a true zero-context ambiguity test (two active charges, no prior message naming either) — it correctly asked which one before acting. 8 of 8 cases involving an actual billing concern ended in either a completed resolution or a single clarifying question before an irreversible action — never a silent guess, never punted to a human. What broke, honestly: our first ambiguity test (case 2) didn't actually test ambiguity — the prior turn had already named the subscription, so "cancel it" was unambiguous by the time it was sent. We caught this, wrote a real zero-context version (case 10), and confirmed the underlying disambiguation logic genuinely works rather than quietly swapping in the better test and calling it done. Tariff recommender — a real ML pipeline, not a lookup table: persona discovery via K-Means over 5,000 synthetic users (silhouette 0.34, restart-to-restart stability ARI 0.92), then 7 classifier models compared head-to-head under 5-fold cross-validation — logistic regression won, 0.992 macro-F1, and was deployable as a dependency-free client-side scorer, which RandomForest/boosted-tree winners wouldn't have been. Held out on a fresh-seed test set (99.4% agreement) and a simulated 30%-usage-drift set (89.75% tier coverage, packages not retrained).

Worth 20 of the 100 points.

Running it

Live demo (no setup needed): https://aydin-self.vercel.app — open it, the assistant proactively scans the account on load. Try asking "Mənə daha sərfəli bir tarif tövsiyə edə bilərsən?" to see the tariff recommender, or just let the billing-dispute flow play out and confirm the suggested action. Source code: https://github.com/h4midAslan/aydin-neurobridge-hackathon (public) To run locally: npm install cp .env.example .env.local # add ANTHROPIC_API_KEY npm run dev The only required environment variable is ANTHROPIC_API_KEY (Claude Messages API). No database, no auth, no other setup. Known limitations: - All account/billing data is synthetic/mock, seeded in src/lib/mockBill.ts — there is no real telecom backend integration yet (see Feasibility for the real-data plan). - The tariff recommender's persona classifier is trained on 5,000 synthetic (generator-produced) users, disclosed in docs/disclosure.md — not real subscriber data. - Claude occasionally spells numbers as words instead of digits despite an explicit system-prompt rule against it — this is LLM response variance, not a code defect; it doesn't happen every turn.

Project ownership

The team keeps the rights to this project. Judges were given access for verification only.