AI Product Categorizer – Alcohol, CNG & CBD

Overview

The AI Product Categorizer is a web-based tool that parses MX item names into clean, export-ready fields such as brand, item name, Kind of Product, size, UOM, container, pack, age, year, and flavor. It supports multiple output formats (Output 1, 1.1, and Output 2 templates) and can run on Alcohol, CNG, CBD, and CAN inputs using the same engine. Batch Mode focuses on throughput and multi-format exports; Review Mode focuses on strict rules, validation, and prompt-level debugging.

My Role

• Designed the full parsing ruleset for Alcohol, CNG, CBD, and CAN domains.
• Implemented the Flask blueprints, GPT-4o-mini integration, and chunked processing logic.
• Built Output 1 / 1.1 / 2 pipelines, including template schemas for different export formats.
• Integrated MySQL logging for user activity, token usage, and Brand Path / Category Path enrichment.
• Created two UIs: Batch Mode (Beast Mode) with multi-output toggles, and Review Mode (God Mode) with a debugging-focused layout.

Tech Stack

• Backend: Python, Flask blueprints, OpenAI GPT-4o-mini, Pandas, MySQL.
• Prompting: external prompt files (Beast & God Mode) and allowed Kind of Product lists.
• Frontend: HTML/CSS, custom JS table renderer, multi-output toggles, dark-mode UI for Review Mode.
• Infrastructure: Linux VPS (Hostinger), Nginx, Gunicorn, file-based logging under /tmp.

Two Modes, Same Engine

Batch Mode (Beast Mode) – High-Throughput Categorizer

• Chunked GPT calls (up to 16 lines per request) with automatic retry and single-line fallback.
• Generates Output 1 (raw GPT), Output 1.1 (enriched with Brand Path & Category Path), and Output 2.
• Dynamic Output 2 template schemas for Alcohol, CNG, CBD, CAN CNG, and CAN Alcohol formats.
• Accent-insensitive item lookups, container normalization, and copy-to-clipboard buttons for all outputs.
• UI lets users switch Output 2 format via dropdown without re-running GPT.

Review Mode (God Mode) – Precision & Debugging

• Uses a stricter prompt with an allowed Kind of Product list to prevent generic or vague KOP values.
• Cleans Item Name by removing brand, size, unit, and container words, then re-normalizing.
• Reuses the same container hallucination guard as Batch Mode, logging any fixes or removals.
• Dark UI focused on reviewing rows, timing, and OIN process counts for debugging runs.
• Logs full prompts, replies, and token usage to help refine rules and manage cost.

Key Features

• Strict schema enforcement: every row outputs exactly 12 fields in a fixed order, ready for downstream Excel / SQL.
• Container hallucination firewall: only sets Bottle/Can/Box/Pouch/Jar/Keg when the term appears as a whole word in the product name.
• L vs ml enforcement: keeps liter sizes exactly as written (e.g., 1.75L → 1.75 + L), never silently converting to milliliters.
• Brand & category enrichment: joins against Brand Path and Category Path data, then builds multiple export layouts from the same GPT output.
• Cost-aware logging: writes user email, action, prompt tokens, completion tokens, total tokens, and cost into MySQL for tracking.

Impact

The dual-mode setup lets teams choose between speed and precision without switching tools. Batch Mode handles day-to-day batch categorization and export, while Review Mode is used for tricky edge cases, rules tuning, and QA investigations. Together, they replace manual spreadsheet parsing and reduce per-line categorization work from minutes to seconds, while keeping outputs structured enough to plug directly into existing Alcohol, CNG, CBD, and CAN workflows.

Demo