# VoiceLedger: Turn a 20-Second Voice Note into a GST Invoice Using Open-Weight Gemma and MongoDB Atlas

> Source: <https://dev.to/atharv96k/voiceledger-turn-a-20-second-voice-note-into-a-gst-invoice-using-open-weight-gemma-and-mongodb-4cfa>
> Published: 2026-10-19 03:45:18+00:00

*This is a submission for the [Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle](https://dev.to/challenges/hacktoberfest-week2-2026-10-12)*

Independent contractors in India—plumbers, electricians, freelance event photographers, mechanics, and home bakers—spend their entire day doing hands-on work. The last thing they want to do at 9:00 PM is sit at a desk, open a spreadsheet, look up HSN/SAC tax codes, separate material costs from labor charges, calculate CGST/SGST splits, and format a PDF bill. Because of this friction, many either delay billing (hurting their cash flow) or send rough text messages on WhatsApp that look unprofessional and make quarterly GST filing a mess.

**VoiceLedger** turns that 45-minute evening chore into a 20-second voice note recorded right after finishing a job.

A contractor simply taps the microphone and speaks naturally in English or Hinglish:

*"Sharma ji ka 2 copper valve change kiya, 1400 parts ka laga, aur 2 ghante fitting labor at my standard 500 rate. 18% GST laga ke bill bana do."*

In under two seconds, VoiceLedger:

`scribe_v1`)` gemma2-9b-it`)` llama-3.3-70b-versatile`)` MATERIALS` - HSN 8481) from service work (`LABOR` - SAC 9954).`upi://pay?...`), 1-click A4 PDF export, WhatsApp share link, and spoken voice confirmation via VoiceLedger is designed as a **₹499/month (~$6/month) or $9/month micro-SaaS** for independent contractors, solo service professionals, and small studios. Getting even a single overdue invoice paid faster—or saving 10 hours of manual bookkeeping a month—pays for the subscription many times over.

*(Tip for Judges: If you are testing in a quiet environment without a microphone, you can click any of the **3 Instant Judge Demo Presets** directly below the voice orb to run the full pipeline immediately.)*

`MediaRecorder` capture (`Ready · Step 1 of 3`), built to handle natural English and Hinglish speech.` OpenAI GPT-OSS 120B`, `Gemma 2`, `Llama 3.3`), a live `12` persisted bills), and real-time `Neha Kulkarni`), configures the applicable tax slab (` 5% GST`), and provides 1-click persistence to `#8188b901994e8bf7`), end-to-end execution latency (` 1017ms`), and a live cost comparison showing `GST Pro`, `Minimal`, `Express`, `Artisan`), inline-editable line items (`₹2,400` + `₹1,080` = `₹3,480` subtotal + `₹174` GST = `₹3,654` Grand Total
VoiceLedger converts unstructured English and Hinglish voice notes into HSN/SAC-compliant GST invoices, automated tax ledgers, and scannable UPI QR payments in under two seconds.

Independent contractors plumbers, electricians, event photographers, mechanics, and home bakers frequently lose billable hours to manual evening paperwork, itemized quoting, and quarterly tax reconciliation.

**VoiceLedger** eliminates manual data entry through a 5-stage open-weight AI agent pipeline. A contractor records a quick voice memo after finishing a job (or selects a pre-configured workflow preset). The backend transcribes the audio via **ElevenLabs Scribe**, retrieves repeat client billing records from **MongoDB Atlas**, extracts a strict JSON schema using open-weight **Google Gemma 2 (`gemma2-9b-it`)** or **Meta Llama 3.3 (` llama-3.3-70b-versatile`)**, verifies line-item arithmetic deterministically, and returns an editable multi-template invoice with a live UPI payment…

VoiceLedger is structured as a decoupled full-stack monorepo (`/client` and `/server`) built around a 5-stage agentic pipeline[cite: 1, 20].

```
[React + Vite Client: MediaRecorder (audio/webm)]
       │
       ▼ POST /api/voice-invoice (multipart/form-data)
[Node.js + Express Backend wrapped in Sentry Root Transaction]
       │
       ├─► Span 1 (ai.stt): ElevenLabs Scribe v1 converts audio buffer to raw text
       ├─► Span 2 (db.query): Mongoose queries MongoDB Atlas for client aliases & rates
       ├─► Span 3 (ai.chat_completions): Open-weight Gemma-2-9b-it / Llama-3.3-70b
       │          extracts lineItems[], category (MATERIALS|LABOR), and HSN/SAC codes
       ├─► Guardrail: Deterministic JS engine computes Subtotal, CGST, SGST & UPI URI
       ├─► Span 4 (ai.tts): ElevenLabs Flash v2.5 synthesizes voice confirmation MP3
       └─► Persistence: Stores finalized document in MongoDB Atlas 'invoices' collection
```

`response_format: { type: "json_object" }`) with low temperature (` 0.1`) for deterministic schema extraction.` eleven_flash_v2_5` generates low-latency spoken confirmations so contractors can verify totals hands-free.
**Data Layer & Client RAG:** **MongoDB Atlas** stores two primary collections (`Client` and `Invoice`). When a transcript arrives, the backend scans the `Client` collection for matching names or aliases (e.g., `"sharma ji"`), injecting the client's saved GSTIN, billing address, and default hourly rate directly into Gemma's system prompt.

**Frontend Experience:** Built with **React 19**, **Vite**, and **Tailwind CSS v4** (`@tailwindcss/vite`). It features inline table editing, real-time SVG QR generation via `qrcode.react`, client-side PDF compilation via `html2pdf.js`, and a mouse-responsive spotlight ledger drawer.

**Agent Observability:** Instrumented with **`@sentry/node`** (` instrument.js`) using custom spans to measure latency and token consumption at every step of the pipeline.

When building a vertical SaaS for price-sensitive independent workers in emerging markets, **unit economics and schema control determine whether the business can survive.**

If a busy contractor generates 150 invoices a month—including client context lookups, multi-item tax classification, and retries—running that pipeline on closed frontier audio/reasoning APIs costs roughly **$0.034 per run** (~$5.10/month in raw API costs per user), wiping out margins once payment gateway fees and hosting are factored in.

By using open-weight **Gemma 2 (`gemma2-9b-it`)** for structured JSON extraction, each inference call consumes ~480 tokens at an effective cost of **$0.00010 per invoice**. Even at 200 invoices per month, the LLM cost per user is under **$0.02/month**, sustaining a **98.8% gross margin**.

| Metric | Open-Weight Stack ( `gemma2-9b-it` ) | Closed Frontier API Stack | 
|---|---|---|
| **Avg. Extraction Latency** | **~390 – 420 ms** | ~1,400 – 2,200 ms | 
| **Cost per Invoice Run** | **$0.00011** | ~$0.03400 | 
| **200 Invoices / Mo Cost** | **$0.022** | ~$6.80 | 
| **Model Portability** | **Swap between Gemma & Llama in 1 click** | Vendor lock-in | 

In the top navigation bar of VoiceLedger, users can switch the underlying extraction engine between `google/gemma-2-9b-it`, `google/gemma-3-27b-it`, and `meta/llama-3.3-70b`. Because our prompt architecture relies on standard open-weight JSON schemas rather than proprietary tool-calling wrappers, we can route requests to whichever open model or inference provider offers the best latency and price-to-performance ratio.

General-purpose closed chatbots try to do everything inside the LLM—including arithmetic, which frequently leads to rounding errors on tax percentages. Our open-weight pipeline restricts the model strictly to **semantic entity extraction** (identifying item descriptions, quantities, unit rates, and HSN/SAC categories) while delegating all tax math and UPI URI formatting to deterministic JavaScript.

`[Optional: Customize this section with your real test experience before publishing, or keep this structure if you tested it with a local contractor/friend!]`

To test whether VoiceLedger actually works outside a developer environment, I put the app in front of `[Insert Name / Role, e.g., a local electrical contractor / freelance photographer friend in Kolhapur]` right after they wrapped up a job.

Instead of typing out a bill at night, they recorded a 15-second voice note in Hinglish listing the parts replaced and their labor charge.

I am submitting **VoiceLedger** for the following partner prize categories:

**Best Use of Gemma:** Uses Google's open-weight `gemma2-9b-it` / `gemma-3-27b-it` as the core structured extraction engine to parse noisy transcripts into line items, classify `MATERIALS` vs. `LABOR`, and assign Indian HSN/SAC tax codes.

**Best Use of ElevenLabs:** Uses **ElevenLabs Scribe (`scribe_v1`)** for accurate Speech-to-Text transcription across English and Hinglish accents, and **ElevenLabs Flash TTS (`eleven_flash_v2_5`)** for hands-free spoken invoice confirmations.

**Best Use of MongoDB Atlas:** Uses **MongoDB Atlas** (`Client` and `Invoice` collections) both for pre-inference RAG context lookup (matching repeat clients and default rates) and for persisting the contractor's tax ledger.

**Best Use of Sentry Agent Tracing:** Wraps the entire Express backend pipeline in a `@sentry/node` root transaction (`voice_to_invoice_agent`) with child spans tracking STT, MongoDB queries, Gemma token usage/latency, and TTS synthesis.

**Best Use of Render:** Deployed the full-stack application on Render.

**Best Use of a .tech Domain:** Hosted live on a custom `.tech` domain.