LAXMIPATI Business Suite
AI-powered payment reconciliation & visual inventory search with DINOv3 embeddings and LLM column mapping.
The Challenge
LAXMIPATI faced two critical bottlenecks: (1) Payment reconciliation chaos - manual Excel processing from Flipkart/Amazon consumed 8+ hours weekly with 15% error rates. (2) Inventory discovery friction - sales staff couldn't find matching products when customers showed reference photos. Text-based search failed for visual attributes like border patterns and weave textures.
The Solution
Developed AI Column Mapper using Claude/Groq to auto-detect platform-specific columns with 95%+ accuracy. Built Visual Search Engine with DINOv3-based 8-crop multi-vector architecture - upload a photo, get top 6 matches in under 3 seconds. Added Editorial Dashboard with warm, magazine-inspired design system.
System Design
Key Results
Key Features
AI-Powered Column Detection
Automatically identifies and maps Excel columns from any e-commerce platform export. Uses Claude Sonnet 4 for semantic understanding of headers like 'Shipping Fee (Collected)' vs 'Reverse Shipping Charges'.
Multi-Vector Visual Search
Upload a reference photo → System removes background, generates 8 strategic crops (full, center, corners), extracts 384D embeddings via DINOv3 ONNX, returns visually similar products ranked by max crop similarity.
Tiered Fallback Search
Vision LLM (Claude) extracts semantic attributes. Search cascades through 6 tiers: Tier 0 (color + product + style + border) to Tier 5 (pure visual similarity).
Real-Time SSE Sync Pipeline
Server-Sent Events stream sync progress as each image is indexed. Per-image commits ensure crash-safe operation — resume from last successful point.
Editorial Design System
Warm, magazine-inspired UI: Fraunces serif headings, terracotta (#c1604a) primary, steel-blue (#5b7c8d) secondary, transparent glass cards, full light/dark mode.
Admin Dashboard & User Management
Role-based access control, activity tracking, user management with role badges. Real-time stats cards, revenue charts, and recent activity feed.
Tech Stack
System Flowcharts
Image Search Pipeline
Multi-vector visual search with 8-crop strategy and tiered fallback
Payment Reconciliation Flow
AI-powered column detection and transaction matching
Multi-Vector Embedding Strategy
8-crop approach captures border patterns and fine details
Deployment Architecture
Screenshots Gallery
Key Learnings
Multi-Vector > Single-Vector for Fashion: Splitting images into 8 strategic crops dramatically improved matching accuracy for sarees with distinctive border work.
DINOv3 Beats CLIP for Pure Visual Matching: Self-supervised learning captures texture and weave information that text-image models miss.
Cookie Auth Complexity on Mobile Safari: ITP blocks third-party cookies. Solution: explicit domain parameter + matching cookie scope.
SSE > WebSockets for Progress Streaming: For one-way updates, SSE is simpler with native reconnection and no handshake overhead.
Editorial Design Creates Trust: Moving from generic SaaS styling to warm editorial palette transformed user perception for financial data handling.