AI Enterprise SolutionsGPT & Claude

FiscalAi

User: the finance / audit team of a mid-size company, and every employee who files an expense. Today the auditor opens the policy, compares each invoice by hand, divides totals by nights and guests, converts dollars and euros, hunts for the approval line, and cannot notice the same invoice arriving twice. The employee sends a PDF and hears nothing until it is paid or bounced. FiscalAI: the employee uploads a PDF or photo and gets an answer in seconds (approved / flagged / needs review) with the broken rule and the exact maths, e.g. "1140 / 3 nights = 380 per night vs limit 300". The auditor sees only what needs a person, with the calculation and the original document side by side, and clears or rejects in one click; the employee is emailed the decision. An assistant answers "may I spend this?" before the money is spent. Design principle: the AI reads, the code decides. A model extracts the fields; a deterministic rules engine makes every compliance decision, so the same invoice always gets the same verdict and text hidden in an invoice cannot talk its way to "approved".

Open the demo Video GitHub

Preliminary score

74.5/ 100Strongaverage of 2 judges

Value for the user21.0 / 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 testing16.5 / 20
Tests, examples of failures, and comparison with the current approach.
Feasibility8.5 / 15
Data requirements, running costs and a clear next step.
Originality6.5 / 10
A meaningful difference from common solutions.

What the judges wrote

Claude Opus 5.5Preliminary73/100
Strengths
Your 'AI reads, code decides' architecture is a clear design choice: a deterministic rules engine plus an approval guard gives you reproducible verdicts, exact maths in every explanation, and a concrete defence against prompt injection and forged approval lines. Your testing is unusually honest and strict. Cases pass only when both the status and the exact set of rule IDs match, and your spreadsheet-filter baseline catches 16 of 29 violations against your 29. You document 24 real failures, each with a fix and a test, including a bug in your own test expectations. The product goes well beyond a demo, with roles, duplicate refusal before the model is called, an audit trail with reopen-with-reason, email notifications, three languages and 77 automated tests.
To improve
Your live-model results are still a placeholder ('__/42 pass, median __ s, $__ per invoice'). That leaves the AI-dependent part unmeasured: free text, Azerbaijani and Russian receipts, scans and injection against the real model. Run it and report the numbers along with the cases that failed. Because the built-in reader handles the sample PDFs without a key, make the demo show clearly where Gemini is essential, such as photos, emails and messy layouts. Add running costs and a concrete next step, for example a pilot with one finance team using real anonymised invoices. Also measure auditor time per invoice today versus with FiscalAI, so you have an adoption metric.
GPT-6 AstraPreliminary76/100
Strengths
Your separation of AI extraction from deterministic policy checks makes the compliance calculations explicit and auditable. Your reported 36/36 offline results use strict verdict-and-rule matching, and the documented failures show meaningful debugging rather than generic success claims. Your employee-to-auditor workflow addresses practical needs through duplicate refusal, decision history, role restrictions and notifications.
To improve
Complete the live-model benchmark: accuracy, latency and per-invoice cost remain blank, leaving the AI extraction path insufficiently evidenced. Compare against actual auditors on realistic receipts, measuring false approvals and review time rather than relying only on a simplified amount filter. Validate approvals against a trusted authorization source, since finding approval text inside an invoice does not authenticate it, and make restricted auditor signup and installation-specific admin credentials the defaults. Define a small finance-team pilot with explicit success thresholds, data-retention controls and estimated monthly operating costs.

The team

Shahriyar HuseynliLead
product
Hasan Hasanov
developer
Rizvan Valiyev
developer

Disclosure

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

Models: Google Gemini (gemini-3.8-flash, via the Gemini API)
Data: All invoices and the company expense policy ("Nordvik Holdings") are synthetic, written for this project. No real company, customer or personal data. With a model configured, the invoice content is sent to that provider's API; with no key nothing leaves the machine. Uploaded files and results are stored locally in SQLite.
Components: FastAPI, Uvicorn, Pydantic, pypdf, SQLite (Python standard library), Google GenAI / Anthropic / OpenAI SDKs, Next.js 14, React 18, TypeScript, Familjen Grotesk and Courier Prime fonts. No public templates beyond framework scaffolding.
Built during the event: Confirmed

Quality testing

What we tested - 42 quality cases with strict scoring: a case passes only if the status AND the exact set of cited rule IDs are right. They cover limit boundaries (150/151, 500, 2000), per-night and per-person maths, the approval hierarchy, USD/EUR conversion, missing data, credit notes, prompt injection and a forged approval line. - Decision logic (offline reader, reproducible): 36/36 pass. The same 36 cases with a spreadsheet-style amount filter, our proxy for the current approach: 14/36, citing 16 of 29 violations vs our 29 of 29. - Live model (gemini-3.8-flash) on all 42 cases incl. 6 free-text ones (email, Azerbaijani receipt, Russian invoice, itemised totals, vendor typo, injection): [__/42 pass, median __ s, $__ per invoice — FILL IN FROM THE RUN] - 77 automated tests: roles and sessions, duplicate refusal, decision history, filters, reports, the assistant (an employee cannot ask about a colleague's invoice), email notification. - A scripted browser run in three languages, both themes and phone width. What broke (24 are documented in QUALITY_TESTING.md; each has a fix and a test) - A note saying "pre-cleared by the Finance Director" could approve a 4800 AZN payment to an unlisted vendor. Fix: code checks every approval claim against the document text. - A 900 AZN flight was judged as a 900 AZN hotel night (wrong category alias). - A refund of -240 was flagged as a 240 AZN meal; "1.234,50" was read as 1.2345. - The same PDF could be submitted twice and both copies approved. Now refused at upload. - A second auditor could overwrite a decision without a trace. Now a decision must be reopened with a reason and every step is kept. - One of our own test expectations was wrong (it missed that a second rule applied); the old runner compared only the status and never noticed.

Worth 20 of the 100 points.

Running it

docker compose up --build 1) Backend (Python 3.11+): cd backend python -m venv venv && venv\Scripts\activate (macOS/Linux: source venv/bin/activate) pip install -r requirements.txt uvicorn main:app --port 8000 2) Frontend (Node 18+), second terminal: cd frontend npm install && npm run dev 3) Open http://localhost:3000/login, click "Set up the admin", then register an employee and an auditor (two browser windows) and upload one of the sample invoices offered on the Submit page. Without a key the built-in reader handles the sample PDFs and every feature works. To read real scans, photos and free text, put LLM_PROVIDER=gemini and GEMINI_API_KEY in backend/.env. Tests: cd backend && pip install -r requirements-dev.txt && python -m pytest -q tests

Project ownership

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