---
sdk: gradio
sdk_version: 6.16.0
app_file: app.py
title: Kirana Detective AI
short_description: AI invoice auditor for kirana stores — find where money is being lost
license: mit
tags:
  - invoice-audit
  - llm
  - yolo
  - gguf
  - gradio
  - indian-fmcg
  - kirana
---

# Kirana Detective AI

AI-powered inventory and invoice auditor for Indian kirana stores. Upload a distributor invoice and delivery photos — Kirana Detective finds pricing anomalies, missing deliveries, duplicate charges, and GST errors, then reports the total rupee leakage with actionable steps.

## Models

All models run **locally — no cloud API calls**.

| Model | HuggingFace Repo | Purpose |
|---|---|---|
| MiniCPM-V 4.6 (transformers) | `build-small-hackathon/minicpm-v-4-6-indian-invoice-extraction-merged` | Invoice OCR + extraction |
| MiniCPM5-1B (GGUF) | `build-small-hackathon/minicpm5-1b-indian-fmcg-normalizer` | Product name normalization |
| YOLO26n (ONNX) | `build-small-hackathon/yolo26n-indian-fmcg-detection` | Delivery photo product counting |

Models are downloaded automatically on first run via `hf_hub_download()` / `AutoModel` and cached locally.

## Running Locally

```bash
pip install -r requirements.txt
python app.py
```

Requires ~6 GB RAM. First run downloads ~3 GB of model weights.

## Six-Agent Pipeline

```
Invoice Upload → Agent 1 (Extract) → Agent 2 (Normalize) → Agent 3 (Pricing)
→ Agent 4 (Count) → Agent 5 (Reconcile) → Agent 6 (Report) → ₹ Savings
```

Each agent run is traced and published to HuggingFace Hub for the Sharing is Caring badge.

## Fine-Tuned On

- **Invoice extraction**: 500 synthetic Indian invoices (printed GST, handwritten, Tally PDF, WhatsApp screenshot) across 10 major FMCG suppliers
- **Product normalization**: 2,000 synthetic `(raw_name, normalized_name)` pairs covering 200 Indian FMCG SKUs
- **Product detection**: Indian Grocery Object Detection dataset (Roboflow) fine-tuned on YOLO26n

Training was done on Modal using Unsloth QLoRA (MiniCPM models) and Ultralytics (YOLO26n).