DEV Community

Arnav Patel
Arnav Patel

Posted on AI-assisted

GhostLedger: The Sovereign Financial Machine — Why Our AI Bookkeeper Math Only Works Because Gemma 2 Is Open-Weight

#hf26challenge #gemma #opensource #ai

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

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

I'm Arnav, from Bengaluru. For Week 1's "Touch Grass" challenge, I built WildTrail to get people outdoors. For Week 2's "Side Hustle" prompt—build something with open-weight models or open-source AI that someone would actually pay for—I wanted to solve a problem that every freelancer, consultant, and indie hacker runs into every tax season: the dreaded receipt shoebox.

Every year, independent contractors leave an estimated $3,000 to $6,000 on the table in legitimate business tax deductions under IRS Schedule C. Manual sorting through flight itineraries, lunch receipts, hardware invoices, and SaaS bills is painful. But the existing solutions have two fatal flaws:

  1. The NDA & Confidentiality Breach: Uploading client receipts to cloud SaaS (Expensify, Dext) or proprietary cloud LLM APIs sends confidential client names, project codenames (Project Apex), secret meeting locations, and credit card snippets straight into third-party databases and external AI training sets.
  2. The Unit Economics Trap: If you want to sell a $29 one-time desktop app or a $9/month micro-SaaS to freelancers, paying $0.03 to $0.08 per multimodal receipt call to closed cloud APIs completely obliterates your gross margin.

So I built GhostLedger // 2100: a 100% air-gapped, zero-leak receipt and expense intelligence tool powered by Google's Gemma 2 open-weight model family and IRS Schedule C tax rules.

🏛️ What I Built & How It Works

GhostLedger operates like a sovereign financial flight deck. It takes raw receipt text, invoice emails, or flight itineraries, and routes them through a 5-stage pipeline:

flowchart TD
    subgraph Ingestion ["1. Multi-Stream Ingestion"]
        R1["📄 Raw Receipt Text"]
        R2["📧 SaaS Invoice Email"]
        R3["🎫 Flight / Hotel Itinerary"]
        R4["💻 Stdin Unix Pipe"]
    end

    subgraph Privacy ["2. Client NDA Shield"]
        PR["🛡️ PrivacyRedactor"]
        PR --> |Sanitizes| M1["Credit Cards / PII"]
        PR --> |Anonymizes| M2["Client NDA Names"]
        PR --> |Masks| M3["Project Codenames"]
    end

    subgraph AI ["3. Open-Weight Inference Engine"]
        GM["🧠 GemmaEngine"]
        Ollama["Local Ollama (gemma2:2b / 9b)"]
        Fallback["Deterministic Rule Engine"]
        GM --> Ollama
        GM -.-> Fallback
    end

    subgraph Tax ["4. IRS Schedule C Intelligence"]
        TC["⚖️ TaxCategorizer"]
        TC --> C1["Meals (50% § 274(n) Line 24b)"]
        TC --> C2["Travel & Lodging (100% Line 24a)"]
        TC --> C3["Software & Cloud (100% Line 18)"]
        TC --> C4["Office & Section 179 (100%)"]
    end

    subgraph Output ["5. Multi-Surface Delivery"]
        EX["📤 ExpenseExporter"]
        EX --> CSV["QuickBooks / Xero CSV"]
        EX --> MD["IRS Audit Markdown Report"]
        EX --> JSON["Clean Accounting JSON"]
        EX --> WEB["🖥️ 3D Command Cockpit"]
        EX --> CLI["📟 Standard Output (stdout)"]
    end

    Ingestion --> Privacy
    Privacy --> AI
    AI --> Tax
    Tax --> Output

1. 🧠 Google Gemma 2 Local Inference

GhostLedger integrates directly with Google's Gemma 2 open-weight model family. When running locally with Ollama (ollama run gemma2:2b), it communicates with http://localhost:11434 to extract merchants, dates, currencies, itemized line items, and tax amounts. If Ollama is offline or the user is in a pure browser/edge environment, GhostLedger automatically triggers a deterministic zero-dependency heuristic parser with zero downtime.

2. 🛡️ Client & NDA Privacy Shield

This was my favorite feature to build. If you have an invoice for a client dinner with Client: Initech discussing Project Apex, handing that raw receipt to an external accountant could violate non-disclosure agreements. GhostLedger's privacy shield runs on-device before export:

  • Identifies project codenames (Project Apex ➔ Project [CONFIDENTIAL_CLIENT_1])
  • Identifies client names (Client: Initech ➔ [REDACTED_NDA_INFO])
  • Scrubs credit card numbers, phone numbers, and private emails

3. ⚖️ IRS Schedule C Tax Audit Defense

GhostLedger doesn't just total numbers—it understands tax law:

  • Meals & Entertainment (IRC § 274(n)): Automatically applies the mandatory 50% limit (Line 24b) and notes that business purpose must be documented.
  • Travel & Lodging (IRC § 162(a)(2)): 100% deductible for travel away from tax home (Line 24a).
  • Software & Cloud Infrastructure: 100% deductible under Line 18 / Other Expenses for ordinary and necessary developer services (DigitalOcean droplets, SaaS subscriptions).
  • Hardware & Office Equipment: 100% deductible in year acquired under Section 179 and the De Minimis Safe Harbor for purchases under $2,500.

4. 📟 The Unix Philosophy

GhostLedger adheres to the Unix philosophy: clean text streams, composability, and quiet modes:

# Pipe receipt text directly into jq
cat samples/saas_subscription.txt | ghostledger -q -f json | jq .total_deductible

# Stream sanitized RFC-4180 CSV with client tokens redacted
ghostledger samples/ -q -r -f csv > accountant_safe_expenses.csv
Enter fullscreen mode Exit fullscreen mode

5. 🌌 Dual-Interface Experience

  • The First-Principles Mission Manifesto: A 3D hyperspace warp drive (450 animated depth-mapped particles), an interactive compounding ROI tax calculator, and a Unix stream holodeck.
  • The Financial Command Cockpit: A high-density workstation with 2-column ingestion, telemetry status, live ledger search filter, IRS breakdown radar, and a collapsible POSIX log drawer.

💰 The Side Hustle Business Math: Why Open Innovation Wins

The prompt asked: Why does open innovation matter for what you built? Could someone actually pay for it?

Here is the exact unit economics breakdown that makes GhostLedger viable as a commercial side hustle:

Expense Tool Architecture Marginal Cost Per 1,000 Receipts Gross Margin on $29 App Data Privacy Level
Proprietary Multimodal Cloud API ($0.05/call) $50.00 Negative or razor-thin ❌ Data sent to cloud
GhostLedger (Open-Weight Gemma 2) $0.00 100% Gross Margin ✅ 100% Local / Air-Gapped

The Monetization Model

  1. $29 One-Time Lifetime License: A local Electron/Desktop app or CLI binary for privacy-conscious software engineers, lawyers, and consultants who refuse cloud SaaS. Zero recurring server costs for me as the indie developer.
  2. $9/Month Sovereign Self-Hosted Team Edition: A Dockerized team container for boutique consulting firms who want shared audit reports without their financial data touching third-party clouds.

Because Google's Gemma 2 runs on Apple Silicon or standard laptop CPUs at zero marginal cost, the business math actually works.

🛠️ What Went Wrong (The Engineering Hurdles)

Just like in Week 1, I made a strict rule: verify every claim and inspect real network and runtime behavior. That rule caught several tricky bugs:

1. The Vercel Dual-Repository Trap

When I used the "Deploy with Vercel" button, Vercel cloned the project to a new repository named XiaoArnav/ghostledger (without the hyphen). For several commits, I was pushing updates to XiaoArnav/ghost-ledger (with a hyphen) and wondering why Vercel was still serving an earlier build! Inspecting the GitHub API revealed the second repo. I synced both and configured Git dual push URLs (git remote set-url --add --push origin ...) so every push now deploys to both repositories simultaneously.

2. Legacy builds vs. Modern Vercel Routing

An early iteration of vercel.json used legacy @vercel/static and @vercel/python builds blocks. On modern Vercel, this caused static routing conflicts and returned 404s for HTML subpages. I removed the deprecated builds blocks and migrated to clean modern cleanUrls: true with rewrites.

3. 3D Particle Text Selection Glitch

On the Mission Manifesto landing page, clicking or dragging on the 3D canvas caused the browser to highlight hero text. I solved this by adding strict user-select: none; across the viewport and suppressing rapid multi-clicks (e.detail > 1).

4. Air-Gapped Client Resilience

To ensure the live web app works even if an edge serverless function experiences a cold-start delay, I built a complete client-side deterministic fallback engine inside public/app.js. If the serverless /api/parse endpoint is delayed, the local engine immediately itemizes the receipt in-browser without the user ever noticing.

🏆 Prize Categories Entered

  • Overall Winner: A practical, real-world utility solving a $3,000+ problem for freelancers with 100% gross margin side-hustle economics.
  • Featured Category: Best Use of Gemma: Employs Google's Gemma 2 open-weight model family for air-gapped structured extraction and tax deduction reasoning without cloud leaks.
  • Featured Category: Best Use of DigitalOcean: Includes automated classification for cloud infrastructure invoices and zero-dependency production deployment blueprints.

🏷️ Hacktoberfest 2026 Tags & Discovery Index

GhostLedger is indexed across official Hacktoberfest, DEV, and GitHub channels using three distinct tag categories:

  1. DEV Challenge Submission Tags:

    • #hf26challenge — Official required challenge tag for the Hacktoberfest 2026 Open-Source AI Challenge series.
    • #gemma — Enters the Featured Partner Category: Best Use of Gemma ($200 prize + exclusive badge).
    • #opensource — Connects with the global open-source developer ecosystem.
    • #ai — Indexes into the open-weight and open-source AI builders feed.

  2. GitHub Repository Topics:

    • hacktoberfest, hacktoberfest2026, side-hustle, gemma, gemma2, privacy, taxes, schedule-c, zero-leak, air-gapped, accounting

  3. Package & Metadata Keywords:

    • Declared in pyproject.toml and setup.py: hacktoberfest, hacktoberfest2026, side-hustle, gemma, gemma2, privacy, receipt-scanner, taxes, schedule-c, accounting, freelance, zero-leak, air-gapped.

🚀 Try It Out

You can also run it locally in 10 seconds (standard Python library only, zero pip installs needed):

git clone https://github.com/XiaoArnav/ghostledger.git
cd ghostledger
python3 -m privatereceipt.cli demo
python3 -m privatereceipt.cli serve --port 8765
Enter fullscreen mode Exit fullscreen mode

Thanks to DEV and the Hacktoberfest team for another incredible challenge! 👻⚖️

Top comments (0)