AI-powered document intake for extracting, validating, and structuring financial data.
Intake AI turns photos and PDFs of receipts, invoices, and other financial documents into structured, editable data that's ready to review and export into spreadsheets, accounting workflows, or databases.
Originally built during HawkHack 2025 at Montclair State University, then expanded into a more complete document-processing app.
Intake AI takes an unstructured financial document and turns it into structured, reviewable, export-ready data.
Manually entering data from receipts and invoices is repetitive, slow, and error-prone. Intake AI explores a faster way to do it.
Users upload an image or PDF, and the app uses Google's Gemini API to pull out fields like:
- Document type, vendor, invoice date, currency, total amount
- SKUs, item descriptions, quantities, unit prices, line totals
- GL categories
The extracted data lands in an editable interface so users can check it over before exporting as CSV, Excel-compatible XLS, JSON, or QuickBooks-compatible CSV.
The goal isn't just reading text off a document — it's turning that unstructured document into structured data that actually fits into a real workflow.
- Image and PDF document processing, including multi-document batches
- AI-powered extraction using Google Gemini
- Structured document and line-item output
- Editable results before export
- Automatic GL category suggestions
- Math mismatch detection and recalculation
- Duplicate document detection
- CSV, Excel-compatible (
.xls), JSON, and QuickBooks-compatible exports - Print / Save as PDF
- Currency conversion previews
- Session history and analytics (spending breakdowns, categories, document types)
- Multi-language interface, light/dark themes
- Demo mode for trying it without burning API usage
Image / PDF
↓
Next.js API Route
↓
Google Gemini
↓
Structured Document Data
↓
Validation & Review
↓
CSV / Excel / QuickBooks / JSON
- User uploads an image or PDF of a financial document.
- It's sent to a server-side Next.js API route.
- Gemini analyzes the document and returns structured data against a defined response schema.
- The app layers on additional logic — validation, duplicate detection, calculations, formatting.
- The extracted data is shown in an editable interface for review.
- The final data exports for spreadsheets, accounting workflows, databases, or wherever it's needed.
Upload — receipts, invoices, and other financial documents as images or PDFs, with batch support.
Extraction & Review — verify fields, review line items, and fix anything before export.
History & Analytics — processed documents stay available for the session, with lightweight analytics on totals, categories, currencies, and recent activity.
JavaScript / JSX · Next.js · React · Tailwind CSS · Google Gemini API · Vercel
Also leans on browser APIs for file handling, clipboard support, printing, and client-side export generation.
This was the first project where a generative AI model became part of the actual data pipeline rather than a chatbot bolted on the side.
Instead of treating Gemini as a conversational layer, the app sends it documents and asks for structured output that the rest of the system can act on directly. Gemini handles the probabilistic part — reading and understanding the document — while ordinary application logic handles the deterministic part: validation, duplicate detection, math checks, recalculation, workflow state, risk flags, and export generation.
Building this gave me real hands-on experience with multimodal AI, structured/schema-constrained model output, prompt design, and designing interfaces around AI output that's usually good but still needs a human check.
Intake AI started at HawkHack 2025 at Montclair State University. I wanted to build something rooted in data and databases while solving an actually annoying problem — the repetitive grind of manually transferring numbers from receipts and invoices into spreadsheets and accounting tools.
The hackathon was a good excuse to experiment with using AI to bridge unstructured documents and structured data. After the event, I kept building — improving the interface, validation flow, export options, analytics, and localization.
- Integrating Google's Gemini API and working with multimodal inputs
- Designing structured, schema-constrained AI responses
- Keeping AI-generated output cleanly separated from deterministic app logic
- Handling API errors and usage limits gracefully
- Building editable UI around AI-generated data
- Generating CSV, Excel-compatible, and JSON exports
- Managing multi-step processing state in React
- Building responsive UI with React and Tailwind, including light/dark themes
A lot of the frontend patterns from this project are ones I plan to reuse going forward.
Accounts payable, bookkeeping, invoice processing, expense management, procurement, financial reconciliation, spreadsheet prep, database entry, document digitization, and general internal business tooling.
Traditional Workflow Intake AI Workflow
Receipt / Invoice Receipt / Invoice
↓ ↓
Manual Data Entry Intake AI
↓ ↓
Spreadsheet / Accounting System Review Structured Data
↓
CSV / Excel / QuickBooks / JSON
git clone https://github.com/smallqui/intake-ai.git
cd intake-ai
npm install
cp .env.example .env.localAdd your Gemini API key to .env.local:
GEMINI_API_KEY=your_key_hereStart the dev server:
npm run devThen open http://localhost:3000.
- AI-generated data can contain errors and should be reviewed before use in real financial systems
- Currency conversion uses static approximate rates, not a live exchange-rate feed
- Sensitive-data detection is lightweight pattern matching, not a compliance-grade solution
- Document history is session-based, not persisted to a database
- QuickBooks export produces compatible CSV data rather than a native IIF file
- Persistent storage and database integration
- User accounts
- Live currency exchange rates
- Additional accounting-platform exports
- More advanced document validation and confidence scoring
- Larger batch-processing workflows and expanded analytics
- Support for more document types