HesabCheck — AI three-way matching that catches overbilling before you pay
Accounts-payable accountants in Azerbaijani SMEs (retail, construction, distribution) who must check every supplier invoice against the purchase order and the warehouse goods receipt before paying. Today this "three-way match" is done by hand: three documents in three formats (the company's Excel/1C order, a phone photo or scan of the receipt, the supplier's own PDF invoice), the same product named differently in each ("A4 kağız 80 q/m²" / "Бумага офисная A4" / "Office A4 80"), often without product codes, in AZ/RU/EN. Short deliveries and price increases hide in long invoices, so companies pay for goods that never arrived, and the accountant spends hours comparing lines and collecting evidence for disputes. HesabCheck reads the three documents in any format with AI, recognises the same product across languages, calculates the overbilled amount to the cent with deterministic code, shows the page and quote behind every number, and drafts the dispute letter. Example: order 100, received 80, invoiced 100 × 12 AZN → "Mismatch — 240.00 AZN" in seconds.
Preliminary score
- Value for the user22.0 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI25.0 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing17.0 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility8.5 / 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 73.0 / 100 on stage.
What the judges wrote
- GPT-6 AstraPreliminary79/100
- Strengths
- Your separation of AI extraction and multilingual matching from deterministic money calculations is well suited to an accounting workflow. Your reproducible scenarios cover mixed file formats, combined-document splitting, human confirmation and dispute drafting, with a concrete 60-line example reporting four detected errors in 33 seconds. Your testing account is unusually specific about failures and fixes, while clearly distinguishing mocked tests from live AI checks and acknowledging synthetic-data limits.
- To improve
- Your next validation should use consented, anonymized company documents and measure extraction accuracy, missed discrepancies, false alarms and review time against a timed manual baseline. Your feasibility case needs measured API cost per document set, storage and retention requirements, and a concrete next pilot step. Before handling real invoices, replace the shared demo account with isolated customer access, protect stored files, and validate poor scans and VAT-bearing cases.
- Claude Opus 5.5Preliminary80/100
- Strengths
- Your problem framing is excellent: a specific user (AP accountants in Azerbaijani SMEs), a concrete pain (multilingual, code-less three-way matching across Excel, scans and PDFs), and a clear outcome in AZN. You split the work sensibly, with AI for extraction and cross-language product matching and deterministic Decimal code for the money, plus page-and-quote evidence and a drafted dispute letter. Your testing section is among the strongest possible: 308 automated tests, live Gemini runs with planted errors found to the cent, and honestly documented failures. Those failures include the invoice 'Note' prompt-injection case and your structural fix for it.
- To improve
- Your comparison with the manual process is qualitative. Time an accountant on the same 60-line set and report minutes and missed errors next to your 33 seconds. Your feasibility story lacks running costs and a concrete next step: give cost per document set with Gemini Flash-Lite, and name a pilot with a real SME archive to measure accuracy on real scans and handwriting. Automating VAT-bearing invoices should be a priority, since most real Azerbaijani B2B invoices carry VAT.
The team
Musa Budaqov
developer
Mətin QardaşovLead
developer
Kamran Hüseynov
developer