Zero Shot
Problem: Accountants in Azerbaijan work across invoices, PDFs, Excel files, tax legislation, accounting systems and news, and check much of it by hand. Generic AI could speed this up, but it can invent numbers, make up legal answers, or take unsafe actions in financial systems, so it can’t be trusted with the work. Who it’s for: Azerbaijani companies, accountants, finance teams and compliance professionals who need AI help without giving up accuracy or control. What LexAudit AI does about it: It sits between the AI and the financial operations. The AI proposes, the backend verifies with deterministic calculations and official legal sources, and humans approve important actions. Everything is recorded in an audit log and kept isolated per company.
Preliminary score
- Value for the user18.0 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI20.0 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing16.0 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility8.5 / 15
Data requirements, running costs and a clear next step.
- Originality6.5 / 10
A meaningful difference from common solutions.
What the judges wrote
- Claude Opus 5.5Preliminary70/100
- Strengths
- Your quality testing is unusually rigorous for a hackathon. You benchmarked E5, TF-IDF and hybrid RRF retrieval on 159 official DVX tax questions with Hit@k, recall and MRR, then showed that the hybrid approach beats semantic-only retrieval. You were honest about what broke (multi-article recall, the comma-separated citation parser bug, irrelevant receipt retrieval, LLM-invented source IDs) and labelled your limitations clearly. Your design principle that the LLM proposes while deterministic Decimal-based rules calculate and humans approve fits the trust problem in Azerbaijani accounting well.
- To improve
- Name and measure the adoption metric your track asks for, such as minutes saved per invoice or the error rate against an accountant's manual check. Without it, the user outcome is only asserted. Many core workflow pieces (VAT return drafts, ledger workflows, Excel imports) are still pending, so show one complete invoice-to-approved-entry run with real numbers. Manually verify a subset of your benchmark labels and test OCR on more than one receipt. Add running-cost estimates, and explain how sending documents to the Gemini API fits your claim of Azerbaijan data residency.
- GPT-6 AstraPreliminary68/100
- Strengths
- Your Azerbaijan-specific workflow addresses a concrete accounting need, with sensible separation between AI suggestions, deterministic monetary checks, and human approval. Your 159-question retrieval comparison demonstrates measurable gains from hybrid retrieval and honestly reports incomplete multi-article coverage. Your README also establishes useful integration boundaries by retaining ambiguous receipts as needs_review evidence rather than inventing invoice fields.
- To improve
- Your next demonstration should trace one invoice from upload through extraction, validation, account-code suggestion, approval, and a logged draft; the materials substantiate individual components more clearly than this complete workflow. Measure accountant review time and correction rate against the current Excel/1C process. Manually verify benchmark labels and expand testing to answer correctness, varied receipts, and irrelevant-document rejection. Specify operating costs and reconcile the deck's Azerbaijan-only processing claim with the Gemini API dependency before handling sensitive company data.
The team
Ayxan Bədəlov
developer
Riad ŞirəliyevLead
data
Eldəniz Arifzadə
data