I Built DukaanFlow AI: A Local-First AI Order Desk for Tiffin Sellers

What I Built

I built DukaanFlow AI, a local-first AI order desk MVP designed for tiffin and home-food sellers who receive customer requests in Hindi, Hinglish, and English.

Small food businesses often handle delivery skips, menu questions, and subscription requests through messages. DukaanFlow AI explores how AI can help organize these requests while keeping the seller in control.

Key features:

• Classifies customer requests into useful categories.
• Supports an optional local Ollama model using Qwen2.5 1.5B.
• Uses conservative rule-based fallback logic.
• Requires seller confirmation before changing order records.
• Treats one-day delivery skips separately from recurring subscription cancellations.

Demo

Live website: https://utkarshydv27.github.io/DukaanFlow-AI/

The website includes an interactive sample-message demo. This demo uses simple browser-side rules and is illustrative; it is not connected to the Python backend or a live AI model.

Code

GitHub repository: https://github.com/utkarshydv27/DukaanFlow-AI

The project is open source. You can explore the code, review the tests, and suggest improvements.

How I Built It

I used Python, Flask, SQLite, and an optional local Ollama setup with Qwen2.5 1.5B.

The MVP follows a simple workflow:

1. A customer request is entered.
2. The request is classified using the available local model or fallback rules.
3. The seller reviews the suggested action.
4. The seller confirms before an order record is changed.

I also created a responsive static website to explain the project and demonstrate sample request classification. The local test suite passed 19 tests during development.

Why Open Innovation Matters

I wanted to explore how open-weight models can make AI experimentation more accessible to small projects and developers.

Local inference can offer more control over where processing happens and may reduce dependence on hosted AI APIs. However, it still has hardware, electricity, and setup costs. It also needs real-world evaluation before we can claim that it performs better than other approaches.

For me, open innovation means being able to inspect, test, adapt, and improve the tools we build.

My Agent Session (Optional)

I used an AI coding assistant during development to help with implementation, debugging, documentation, and project organization. I reviewed the changes and tested the application during development.

My First User (Optional)

This is an early MVP. I have not yet validated it with a real food seller, so I am not claiming user feedback or real-world adoption.

Prize Categories

DukaanFlow AI explores local-first AI and open-weight models through a practical workflow for small food businesses. Its focus is on accessible experimentation, transparent limitations, and human confirmation before consequential order changes.

The project is still an MVP. WhatsApp Cloud API integration is not live, and real-world seller testing remains future work.

Built by Utkarsh Yadav