KirAI - Financial Report Analysis

Dekstop UI
Mobile UI

Inspiration

Finance teams, especially SMEs, often spend too much time manually reviewing PDF reports and spreadsheets. Important risks can be missed because the data is scattered across pages, tables, and different reporting periods.

We built KirAI to make financial review faster, clearer, and more trustworthy. The goal is not just to summarize a report, but to explain what needs attention with evidence from the original document.

What It Does

KirAI lets users upload a financial report in PDF, CSV, or Excel format. The system analyzes the report and generates:

A plain-language financial summary
Key trends such as revenue growth or expense movement
Risks and unusual patterns
Practical recommendations
Revenue and expense charts
A financial health score
Clickable evidence links back to the source document

If a second document is uploaded, KirAI can also compare two periods, such as Q1 versus Q2.

How We Built It

KirAI is built as a full-stack web app.

The frontend uses Next.js, React, Tailwind CSS, Motion, Recharts, and react-pdf. It provides a simple upload flow, a results dashboard, charts, badges, and a PDF evidence viewer.

The backend uses Node.js, Express, multer, pdfjs-dist, xlsx, and Groq AI. The backend pipeline works like this:

Upload a PDF, CSV, or Excel report.
Extract text from the document page by page.
Mask personal information such as emails, phone numbers, IDs, names, and account numbers.
Calculate financial metrics in code.
Send only masked text, calculated metrics, and optional user instructions to the AI.
Validate the AI response before showing it to the user.
Display the results with evidence links.

A key design decision was to separate calculation from interpretation. KirAI calculates figures such as:

$$ \text{Net Profit} = \text{Revenue} - \text{Expenses} $$

$$ \text{Net Margin} = \frac{\text{Net Profit}}{\text{Revenue}} \times 100 $$

The AI does not invent or recompute these numbers. It explains the calculated results in clear business language.

What Makes It Different

Most AI tools behave like chatbots. KirAI is designed as an evidence-grounded financial review assistant.

Every major insight can include:

A page reference
A short source quote
A confidence label
A clickable evidence link

This means users can verify where an insight came from instead of blindly trusting an AI answer.

KirAI also includes privacy-focused processing. Personal data is masked before AI analysis, and uploaded files are processed for the current session without persistent backend storage.

Challenges We Faced

One challenge was making AI output reliable. AI responses can be inconsistent, so we designed a strict JSON response format and added backend validation. If the AI returns malformed or weak output, the system can fall back to calculated analysis.

Another challenge was evidence linking. It was not enough to show a summary; we wanted users to trace insights back to the original report. To solve this, the backend keeps page-tagged text and validates source quotes, while the frontend opens a PDF viewer and jumps to the relevant page.

Financial parsing was also challenging because reports can use different labels for revenue, sales, income, expenses, payroll, marketing, and other line items. We handled this with keyword-based extraction and calculation logic suitable for the MVP.

What We Learned

We learned that building useful AI products requires more than calling an AI API. The most important parts are structure, validation, privacy, and user trust.

We also learned that AI works best when it is given a clear role. In KirAI, code handles the calculations, while AI handles interpretation and explanation. This makes the output more reliable and easier to explain during evaluation.