KiranaAI — Visual Billing Counter

KiranaAI is an offline-first visual billing system for Indian kirana and retail stores. Instead of relying on barcode scanners or manual billing, the system uses multimodal AI to identify products directly from photos and automatically add them to the bill. The key challenge we addressed is that most AI product recognition systems fail when they encounter new or local products that were not part of training data. KiranaAI solves this using a hybrid architecture combining Gemma 4 and SigLIP. Gemma 4 handles known catalog products quickly, while SigLIP acts as a visual memory system powered by Qdrant vector search. If Gemma cannot identify a product, SigLIP searches a custom catalog built by the shopkeeper using real product photos. The system supports three intelligent flows: 1. Auto billing for known products 2. Visual fallback search for unknown products 3. Conflict detection when the LLM predicts the wrong SKU A major feature is the conflict detection pipeline. If Gemma predicts one product but SigLIP finds a different high-confidence visual match with no meaningful name overlap, the system pauses and asks the user to confirm instead of silently billing the wrong item. Users can also add completely new products using multiple photos captured from different angles. These images are embedded instantly into Qdrant, allowing the system to learn new products without retraining. The entire stack runs locally using: - Gemma 4 via Ollama - SigLIP embeddings - Qdrant vector database - FastAPI backend - Vanilla HTML/CSS/JS frontend This makes KiranaAI privacy-friendly, cost-efficient, and fully offline-capable after setup. The project demonstrates how multimodal AI and vector search can create practical retail automation systems for emerging markets without requiring expensive hardware or cloud infrastructure.