Nxord

Overview

Nxord is a full-stack B2B web platform that automates order management for food & beverage distributors operating in the HoReCa (Hotel, Restaurant, Catering) sector. Built to eliminate manual order entry, the system accepts orders in any format — PDFs, WhatsApp photos, emails, audio recordings — and uses multimodal AI to turn them into validated, structured records ready for export to the ERP.

The platform is in active production use by a wholesale distributor, processing daily orders with measurable time savings per transaction.

The Problem

HoReCa distributors receive orders through dozens of unstructured channels: handwritten receipts photographed on WhatsApp, scanned faxes, voice memos, emails with inconsistent formatting. Beyond the transcription burden, the data was completely fragmented — each order lived in a separate thread or inbox with no unified record. No client purchase history, no visibility into order trends, no basis for business decisions. Operators had to manually transcribe every item, cross-reference the product catalog, and re-enter everything into the AS400 ERP. The process was slow, error-prone, and scaled poorly with volume.

Solution

Nxord addresses the fragmentation problem end-to-end — from ingestion to intelligence:

1. Multimodal extraction — operators drag-and-drop any format (PDF, photo, audio, email). Google Gemini extracts structured product and quantity data regardless of source, consolidating fragmented channels into a single pipeline.
2. 3-stage semantic matching — exact code lookup → fuzzy string matching → cosine similarity over pgvector embeddings. Handles spelling variations, short codes, inconsistent naming across sources.
3. Commercial Intelligence — once orders are structured and unified, the accumulated data powers a full analytics layer: per-client purchase history, YoY trends, basket-analysis upsell opportunities, global KPI dashboard with Pareto concentration, and on-demand AI briefs streamed word-by-word via SSE.
4. ERP integration — validated orders exported in AS400 fixed-width format via SFTP, with full audit trail per order.

Key Features

Multimodal AI Extraction

• Accepts PDF, images (JPG, PNG), audio, and plain text
• Google Gemini processes files asynchronously via Celery workers
• Handles WhatsApp photos of receipts, scanned faxes, informal messages

3-Stage Semantic Product Matching

• Stage 1: exact code lookup
• Stage 2: fuzzy string matching
• Stage 3: cosine similarity over pgvector embeddings
• Confidence score per item; uncertain matches flagged for review

Architecture Decisions

Offline vs online computation. The analytics pipeline is split: a batch job processes the sales history CSV in pure Python and stores pre-computed JSON reports in PostgreSQL. At request time, the API does a single SELECT — no computation, no ORM joins.

pgvector for semantic matching. Every product name is embedded at upload time. Order items are matched by cosine similarity. The 3-stage pipeline (exact → fuzzy → semantic) keeps precision high.

Session auth, no JWT. Django session cookies with credentials: include. The right choice for a B2B internal tool where server-side invalidation matters.

SSE streaming for AI briefs. Gemini takes 5–15s to produce a brief. Server-Sent Events stream the response word-by-word to eliminate the perception of waiting.