This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle
What I Built
Every small shop has a familiar challenge: keeping track of who owes what.
A customer buys groceries today and promises to pay tomorrow. Another customer pays back part of their outstanding balance. Someone else has credit from last week. Traditionally, many small businesses manage these transactions in a notebook called a bahi-khata.
I wanted to explore how AI could make this everyday bookkeeping process easier.
That's why I built TalkLedger — a voice-first digital udhaar ledger designed for small businesses.
TalkLedger brings the familiar idea of a shopkeeper's account book into a digital interface, with conversational AI and a workflow designed around reviewing transactions before saving them.
What can TalkLedger help with?
- Voice-first bookkeeping: Make transaction entry more natural through speech-based interaction.
- Customer credit tracking: Organize customer records and outstanding balances.
- Payment management: Keep track of credit transactions and repayments.
- Gemma-powered conversational AI: Explore how an open-weight language model can interpret bookkeeping requests.
- Review before saving: Let the merchant review, edit, confirm, or discard proposed transaction details.
- Dashboard and reports: Bring ledger information into one digital workspace.
- Repayment forecasting: Explore how historical transaction data could support repayment predictions.
Demo
🎥 Watch the TalkLedger demo video:
https://www.youtube.com/watch?v=xJLpCL9pPh8
🌐 Try the live application:
https://talkledger-4.onrender.com/
Code
https://github.com/Nitu-Bathyal/TalkLedger
TalkLedger is an open-source project built with the intention of making its implementation accessible to other developers.
You can explore the code, understand the application architecture, and contribute ideas for improving the bookkeeping workflow, AI integration, database design, and overall usability.
How I Built It
I built TalkLedger as a full-stack application rather than just a standalone chatbot.
1. Frontend and user experience
The frontend uses React, Vite, and TypeScript.
The design takes inspiration from traditional Indian account books, with warm cream backgrounds, green accents, ledger-style layouts, and a friendly interface.
The application includes dedicated areas for the home page, dashboard, Gemma AI assistant, voice interaction, ledger, payments, forecasting, and reports.
I wanted the interface to feel approachable for small-business owners rather than like complicated accounting software.
2. Backend and database
The backend uses Node.js, Express, and TypeScript, with Prisma ORM for database access.
The deployment setup uses PostgreSQL through Neon, while Render hosts the application.
This architecture connects the user interface with backend routes and database operations, providing the foundation for managing customer and transaction information.
Moving from local development to cloud deployment also gave me practical experience with environment variables, database connectivity, API routes, and deployment configuration.
3. Exploring Gemma 2 for conversational bookkeeping
A key part of TalkLedger is Gemma 2, Google's open-weight language model family.
I designed the conversational assistant around a specific task: helping a shopkeeper express bookkeeping requests in natural language and preparing structured actions for review.
The intended workflow is:
A merchant enters a request through speech or text.
The request is sent to the backend.
The configured model provider processes the request.
The application prepares the proposed transaction details.
The merchant reviews and edits the details if needed.
The merchant confirms the action before it is submitted for saving.
The architecture explores an extensible provider approach, including integration paths for Groq, Hugging Face, and local Ollama inference.
4. Designing for user control
A ledger deals with financial records, so an AI-generated suggestion should not automatically be treated as an authorized transaction.
TalkLedger's proposed-action workflow is designed to separate interpreting a request from confirming a database change.
The merchant can review the customer, amount, and transaction type before confirming. If the action is not confirmed, it should not be written to the ledger.
This design principle is important to me: AI should assist the merchant, while the merchant remains in control of financial actions.
5. Deployment with Render
I deployed TalkLedger on Render, with PostgreSQL configured through Neon.
Render provides the hosting environment for making the application accessible through a public URL.
This part of the project helped me understand that building a web application involves more than designing a good interface. Backend configuration, database connectivity, environment variables, and production testing are also essential.
Why Does Open Innovation Matter?
I wanted to work on a practical problem that small businesses face every day, rather than building another general-purpose AI chatbot.
Open innovation makes it possible to experiment with open-weight models, inspect application code, adapt integrations, and learn from other developers.
For TalkLedger, this creates opportunities to explore how an open-weight model such as Gemma can support a specific workflow, how different model providers can be used, and how developers can improve the application without making every design decision dependent on a single proprietary service.
It also allows other developers to examine the transaction workflow and suggest improvements to reliability, usability, and user control.
My broader goal is to explore how open-source AI can make everyday digital tools more accessible to small businesses.
My Agent Session
I used AI-assisted development tools while working on the application structure, interface, and integration workflow.
I have not included a public DevRelay or agent-session link in this submission.
My First User
My first feedback came from my father, whom I asked to try TalkLedger and share his honest opinion.
It reminded me that building a useful application is not just about adding features. The experience needs to make sense to the people who might actually use it.
This was early feedback on the project, not a paid customer transaction or evidence of regular business use.
Prize Categories
I am submitting TalkLedger to the Hacktoberfest Open-Source AI Challenge — Week 2: Side Hustle.
My intended partner prize categories are:
1. Best Use of Gemma
TalkLedger explores Google's Gemma 2 open-weight model for conversational bookkeeping. The project focuses on interpreting natural-language transaction requests and preparing actions for merchant review.
2. Best Use of Render
TalkLedger is deployed on Render and has a public application URL. Render provides the hosting environment for the project's web application.
Final Thoughts
TalkLedger started with a simple question:
What if keeping track of udhaar(credit) could feel as natural as speaking to someone, instead of manually updating a notebook every time?
Building this project has helped me explore full-stack development, open-weight AI, database integration, deployment, and the importance of designing AI-assisted actions with user control in mind.
There is still room to improve the application, but I am excited about the potential of combining familiar bookkeeping practices with accessible technology.
If you are interested in open-source AI, small-business tools, or making everyday workflows simpler, I would love your feedback and contributions.
Try TalkLedger: https://talkledger-4.onrender.com/
Explore the code: https://github.com/Nitu-Bathyal/TalkLedger
Watch the demo: https://www.youtube.com/watch?v=xJLpCL9pPh8
Top comments (1)
Amazing and Creative...💫⭐