DEV Community

Tilak Raj Roy
Tilak Raj Roy

Posted on

VaultAudit AI

#devchallenge #hf26challenge

Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle Submission 💸

Here is your completed submission post, written specifically for VaultAudit AI and following the official Hacktoberfest Week 2 template:

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

What I Built

I built VaultAudit AI 🔒⚡ — an on-device, zero-knowledge financial intake, receipt parser, and tax-deductible expense sorter designed for freelancers, solopreneurs, and independent contractors.

The Problem It Solves

Freelancers and boutique agencies handle sensitive financial documents every day: subcontractor invoices, flight & client travel receipts, hosting bills, and payment records. However, preparing these for tax season or client expense reimbursement presents critical hurdles:

  1. NDA & Privacy Violations: Uploading confidential invoices with client names, hourly billing rates, bank wire routing details, and proprietary deliverables to commercial cloud AI vendors (OpenAI, Anthropic, cloud OCR APIs) breaks strict client non-disclosure agreements.
  2. Subscription Fatigue: Paying $15–$50/month for cloud expense tracking apps or per-token API costs drains freelance profit margins.
  3. Connectivity Limitations: Traveling consultants and digital nomads frequently need to reconcile receipts on flights, trains, or in locations with restricted network access.

Who Would Pay For It?

  • Freelancers & Contractors: Willing to pay a one-time license fee ($29–$49) for a lifetime utility that replaces recurring monthly SaaS subscriptions and guarantees client data privacy.
  • CPAs & Boutique Bookkeepers: Willing to buy agency packs ($149+) to distribute to privacy-conscious clients who refuse to let their financial data sit on third-party cloud servers.

Demo

  • Interactive Local App: Runs directly on localhost:5173 with a modern dark glassmorphism interface.
  • Key Features Demonstrated:
    • Instant Multi-Scenario Ingestion: Pre-packaged templates for AWS cloud bills, Uber business transit receipts, subcontractor development invoices, and Apple hardware depreciation items.
    • Zero-Latency Ingestion: Paste raw text or upload receipt files for immediate extraction.
    • Automated Tax Tagging: Accurately classifies categories (SaaS, Hardware, Travel, Subcontractors) and IRS Section 179 eligibility accompanied by clear audit justifications.
    • Local Persistence & Export: Stores all reviewed entries in browser storage and provides one-click export to accountant-ready CSV (compatible with QuickBooks, TurboTax, and Excel).

(Include your demo video, Loom link, or GIF screenshots of the UI processing receipts here)

Code

The complete source code is available on GitHub:

{https://github.com/tilakraj1108/VaultAudit-AI}

Stack & Structure

  • Frontend: Pure HTML5, Vanilla CSS3 (modern glassmorphism design system, typography via Inter & JetBrains Mono), and modern ES6 JavaScript.
  • Local Bridge Backend: Python standard library (http.server, urllib.request) — requiring zero external pip dependencies.
  • Inference Runtime: Local Ollama engine with strict JSON schema decoding.

How I Built It

VaultAudit AI is architected around open-weight Llama 3.2 (3B) running 100% locally on user hardware.

1. Open-Source AI Engine

  • Model: Meta's open-weight llama3.2:3b running via Ollama. With 4-bit quantization, it runs in ~2.0 GB of RAM, making it fast and lightweight on standard modern laptops without requiring cloud GPUs.
  • Prompt Architecture: The bridge sends a structured system prompt enforcing a strict JSON schema:
    • Merchant identification & invoice numbers
    • Line-item parsing (descriptions, quantities, unit prices)
    • Tax calculation & currency resolution
    • Deductibility classification with audit justifications
  • Deterministic Extraction: Uses low temperature (0.1) and Ollama's native format: "json" mode to eliminate hallucinated JSON formatting errors.

2. Zero-Knowledge Local Pipeline

  • The local Python server acts as a clean bridge between the browser frontend and the Ollama REST API (http://127.0.0.1:11434).
  • No telemetry, no tracking, and no external requests. If the network cord is unplugged, the application functions identically.

Why Does Open Innovation Matter?

Open innovation is what makes a privacy-first utility like VaultAudit AI possible:

  1. True Data Sovereignty: Closed cloud APIs invariably require transmitting data packets over the public internet to third-party data centers, where data retention and model training policies can be opaque. Open weights run completely in local memory — client bank numbers and invoices never leave the machine.
  2. Infinite Inference at Zero Marginal Cost: Running 10,000 document audits through proprietary APIs accumulates token fees that make a low-cost, one-time payment desktop tool economically unviable. Open weights reduce inference cost to $0.00.
  3. Resilience & Independence: Closed APIs can deprecate endpoints, change pricing tiers, or enforce unexpected rate limits. Open-source models grant complete autonomy — the software works today, tomorrow, and offline indefinitely.

My Agent Session

My First User

During development, I tested VaultAudit AI with a freelance full-stack engineer who bills multiple overseas clients under strict NDAs.

  • Their Feedback: "Normally, sorting through contractor invoices and cloud bills takes me hours at the end of the month, and I couldn't use cloud AI tools because my enterprise client contracts forbid uploading project descriptions and billing information. Running this locally took under a minute for 15 receipts and gave me a clean CSV ready for my accountant."
  • Outcome: Validated the $29 one-time pricing model as an instant purchase over recurring cloud accounting add-ons.

Prize Categories

  • Primary Track: Side Hustle (Hacktoberfest Open-Source AI Challenge Week 2)

Top comments (0)