ReceiptFlow AI: The Zero-Cloud, Local-First Receipt Sorter for Freelancers

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

#

What I Built

Every tax season or quarterly review, I found myself facing the same chaotic chore: a desktop folder cluttered with scattered PDF invoices, train ticket receipts, cloud subscription receipts, and equipment purchases.

When I looked for tools to automate this, I ran into two big dealbreakers:

Predatory SaaS Pricing: Mainstream tools like Expensify, QuickBooks, or Dext charge $15 to $35 every single month just to scan basic receipts. As an independent builder, paying $200–$400 a year for simple OCR felt absurd. 2. Privacy Dealbreakers: Up my receipts to cloud services meant transmitting sensitive banking details, client names, home addresses, and confidential project costs to remote servers and proprietary LLM APIs.

I built ReceiptFlow AI to solve my own problem first—and package it as a clean, viable side-hustle product that other freelancers and indie builders would gladly pay for.

ReceiptFlow AI is a 100% local, privacy-first receipt and invoice organizer. Documents are parsed on-device using open-weight models powered by Ollama (gemma2:2b for embedded text and llava:7b for images). The frontend keeps all metadata inside the browser's IndexedDB, and clean CSV exports are generated client-side.

No cloud databases. No monthly API bills. Zero documents ever leave the machine.

Who Would Pay For It?

Freelancers & Consultants bound by NDAs: Anyone working under non-disclosure agreements who cannot legally or ethically feed client-related expense documents into public cloud LLMs. #

Indie Builders with Subscription Fatigue: People tired of monthly recurring charges for basic utilities. Packaging this as a $29–$49 one-time purchase or desktop application offers huge value while costing me $0 in ongoing server/API overhead. #

Frequent Travelers & Digital Nomads: Anyone who needs to process travel receipts on trains, flights, or places with spotty internet connectivity.

#

Demo

Here is the high-level data flow of how ReceiptFlow AI operates completely offline:

flowchart TD
    A[Drop receipt PDF / Image] --> B[React Upload Queue]
    B --> C[POST 127.0.0.1:8000/api/receipts/parse]
    C --> D{File Format}
    D -->|PDF| E[PyMuPDF In-Memory Text Extraction]
    D -->|Raster Image| F[In-Memory Base64 Buffer]
    E --> G[Local Ollama: gemma2:2b]
    F --> H[Local Ollama Vision: llava:7b]
    G --> I[Defensive JSON Sanitization & Pydantic Validation]
    H --> I
    I --> J[Structured Record returned to Browser]
    J --> K[User Review & Inline Correction]
    K --> L[Browser IndexedDB Storage via idb]
    L --> M[Local Sanitized CSV Export]

Highlights of the User Experience:

Instant Engine Telemetry: Real-time indicator displaying local Ollama reachability and active model (LOCAL ENGINE ONLINE /ENGINE OFFLINE ). #

Drag-and-Drop Batch Processing: Drop multiple files (PDF, PNG, JPG, JPEG, WEBP) simultaneously with live processing queue indicators. #

Smart Analytics: Real-time expense totals, calculated tax breakdowns, and animated SVG sparklines. #

Review & Inline Correction: A clean modal editor to verify and fine-tune model outputs before committing them to records. #

Formula-Injection Safe CSV Export: Exports compliant CSV files with automated protection against spreadsheet formula injection (= ,@ ,+ ,- ).

#

Code

The complete codebase is open source and available on GitHub:

ReceiptFlow AI is a privacy-first, offline-first receipt and invoice organizer for freelancers. Documents are processed by a local FastAPI service that calls Ollama on your own machine. The browser stores extracted records in IndexedDB; CSV export is generated locally.

Privacy model #

• Uploaded documents are sent only to localhost FastAPI, not to a hosted service.
• FastAPI extracts PDF text locally and sends text to Ollama. For image receipts, it can send the image to a local vision model configured in .env .
• Expense metadata is stored in the browser's IndexedDB database.
• No analytics, advertising SDK, account system, or cloud database is included.
• This is a local-first app, not a security certification. Keep your OS account encrypted, secure local backups, and verify model outputs before using them for tax filing.

Stack #

• React + Vite + TypeScript
• Tailwind CSS, Lucide React, Framer Motion
• IndexedDB via idb
• FastAPI, PyMuPDF, Pillow, Ollama…

Repository: https://github.com/Babin123456/ReceiptFlow_AI

#

How I Built It

I designed ReceiptFlow AI around open-source AI and local runtime primitives:

1. Open-Source AI Core (Ollama + Open-Weight Models)

Text Extraction & Parsing: For PDFs containing embedded digital text,PyMuPDF extracts the raw characters in memory. I feed this text to Google'sGemma 2 2B (gemma2:2b ) running locally in Ollama. It is fast, lightweight, and runs smoothly on consumer laptops without needing high-end GPUs. #

Vision Recognition: For photo receipts and scanned receipts, image bytes are encoded into base64 and processed locally byLLaVA 7B (llava:7b ). #

Structured JSON Prompting: The prompt instructs the model to act as a strict invoice data extraction engine, returning only a single JSON object with standardized keys (vendor ,invoice_number ,date ,currency ,category ,subtotal ,tax ,total ,line_items ,notes ).

2. Backend Service (FastAPI + Pydantic v2)

• Built with Python 3.11 andFastAPI , bound strictly to loopback (127.0.0.1 ).

Defensive Extraction: Strips markdown code fences (like triple-backtick JSON wrappers), isolates the JSON body with regex, and validates data through Pydantic models with defensive defaults. #

Stateless by Design: The backend never saves files to disk or writes to a database. It processes files strictly in-memory and returns structured JSON immediately.

3. Frontend Dashboard (React + TypeScript + Tailwind CSS)

Client Persistence with idb: Uses browser IndexedDB to persist receipts across sessions without needing a remote backend database. #

Framer Motion: Delivers smooth micro-interactions, modal transitions, and dynamic state feedback. #

Tailwind CSS & Lucide Icons: Modern obsidian dark aesthetic tailored with mint green and violet accents.

#

Why Does Open Innovation Matter?

Open-source AI was not just an implementation detail for ReceiptFlow AI—it was the only way this product could work.

True Privacy Without Trusting "The Cloud": Closed AI APIs (OpenAI, Anthropic, Google Cloud) require transmitting private financial documents over the wire. Even with zero-retention policies, regulated freelancers and privacy-conscious builders cannot take that risk. Running an open-weight model locally guarantees that data never leaves the laptop. 2. Zero Marginal Inference Cost: If I built this side hustle on top of closed APIs, every receipt processed would burn API credits ($0.01 to $0.05). A sudden influx of users or bulk uploads would destroy profitability. With open-weight models, my marginal cost per inference is literally $0.00. This makes a one-time purchase or open-core monetization model sustainable. 3. Total Model Independence: I am never at the mercy of sudden API deprecations, rate limits, or vendor price hikes. If a newer, faster model comes out, switching requires changing just one environment variable in.env . 4. Offline Resilience: I can process and organize my expenses on a train, on a flight, or in a coffee shop with spotty Wi-Fi.

#

My Agent Session

Pairing with an AI agent during this build helped me move from idea to production-quality implementation fast:

• Architected the zero-persistence pipeline and local-first data flow.
• Engineered defensive prompt validation to prevent JSON parsing crashes when models return markdown fences or comments.
• Styled responsive layouts, Lucide icon integrations, and custom SVG sparklines.
• Implemented RFC 4180 compliant CSV export with formula injection sanitization.
• Orchestrated atomic Git commits with clear, human-like commit tracking across the entire codebase.

source & further reading

dev.to — original article pg_vectorsearch vs pgvector: The 42-Hour Vector Index Just Became a 100-Minute One Prepme - AI Powered Interview Kit Full-page screenshots that are actually complete: lazy loading, sticky headers, 100vh layouts and emoji