Inspiration

If you visit any wholesale market in India, you'll see the exact same scene: a distributor's desk piled with handwritten slips, a clerk with a phone glued to their ear listening to chaotic WhatsApp voice notes in Hinglish, and a desktop running Tally.

Small retailers won't download complex ordering apps—they just want to send a 5-second voice note or a quick photo of an order pad. But for the distributor, processing 50+ of these messy orders every morning takes 3 to 4 hours of tedious typing, leading to avoidable mistakes and wrong deliveries. We wanted to solve this without forcing shopkeepers to change how they work.

What it does

OrderFlow AI turns messy WhatsApp orders into verified, ready-to-bill ERP sales drafts in seconds:

Understands Any Input: Ingests regional voice notes, photos of scribbled slips, and fragmented chat messages.
Smart Catalog Matching: Uses AI to figure out that "Dabur Lal chota" or "Surf Exc 1kg" maps to the exact SKU in the distributor's inventory.
Checks Stock Automatically: Verifies available warehouse inventory before creating the order.
Keeps Humans in the Loop: Shows the clerk a clean side-by-side dashboard comparing the original message with the draft order for a quick 1-click approval.
Exports to Tally: Generates clean Tally XML so the order enters the accounting system instantly, then sends a confirmation receipt back to the retailer.

How we built it

Streamlit for a lightweight, intuitive clerk review dashboard.
Twilio & WhatsApp APIs to receive incoming audio, pictures, and text.
OpenAI Whisper tuned for Indian accents and regional trade jargon.
GPT-4o Vision to read faint handwriting and crumpled carbon-copy slips.
RapidFuzz & Vector Search to bridge colloquial nicknames with distributor SKUs.
SQLite & Python XML tools to handle inventory checks and generate Tally-compatible vouchers.

Challenges we ran into

The hardest part was pack sizes. When someone writes "Dolo 650 2 pkt", do they mean 2 strips, 2 boxes, or a master carton? We solved this by structuring our data models to flag ambiguous units and show the clerk smart suggestions instead of guessing blindly. Handling faint handwriting on crumpled order pads was another hurdle that took careful vision prompt tuning.

Accomplishments that we're proud of

Cutting order preparation time from ~8 minutes of manual typing down to under 20 seconds.
Zero learning curve for retailers—they keep using WhatsApp like they always have.
A practical human-in-the-loop workflow that prevents accounting errors before anything touches Tally.

What we learned

Building for real-world enterprise accounting taught us that confidence scoring matters more than raw automation. It’s always better for an agent to flag an ambiguous item for a human than to confidently guess the wrong product.

What's next

We plan to pilot the prototype with 5 local distributors using their real catalogs, build direct bi-directional sync for Tally and Marg ERP, and add automated WhatsApp voice responses for out-of-stock items.