Automation 2.0 Product Reviewer
The design station of a production pipeline running about 300 personalized orders a day. It hands the designers their hours back.
Internal ToolAI Pipeline
- Stack
- Python, FastAPI, SQLAlchemy, LLM Triage, SVG Rendering
- Status
- In production
- Built at
- Gem Awards, the company where I work
- Screenshots
- Demo data
The Problem
Every Tag Was Designed by Hand, From a Stack of Paper
Before this system, the process started with a printer. Someone printed physical order forms, then highlighted and hand-sorted every product, 40 to 60 minutes before design work could even begin. Then a designer pulled each order up in Amazon Seller Central, copied and pasted every element into Illustrator one at a time, and made manual design edits to every single tag. On busy days that was up to 4 hours per designer of pure transcription and layout, repeated tomorrow, and the day after that.
The Solution
The Pipeline Transcribes. The Designer Designs.
Automation 2.0 killed the paper system outright. It parses each order with an LLM triage layer (Ollama and Claude, with a shadow-mode verifier that catches the AI when it is wrong), auto-sorts the day's run, does the copy-paste for you, and renders a print-ready tag from a versioned JSON document model. It also makes the design calls. AI layout matching reads the customer's request, then places the logo, sizes the name, and sets the colors on its own. The Product Reviewer is where the designers now spend their time, reviewing and art-directing instead of transcribing. Risky orders arrive pre-flagged, a guided mode walks the queue in print order, and a full layer-based canvas editor handles edits in the browser. Standard tags never touch Illustrator anymore, which is why the Adobe licensing bill is coming down.
Under the Hood
It runs on FastAPI and SQLAlchemy behind a session-authed web UI, with a pure-functional SVG rendering core covered by golden-snapshot tests, self-hosted behind Tailscale. The screenshots below show the app running on demo data.
Visual Walkthrough
Open full size
Open full size
Open full size
Key Features
What It Does
- LLM order parsing with a shadow-mode verifier that audits the AI's work
- Auto-sorting of the day's run, replacing 40 to 60 minutes of paper sorting
- Logos, names, and titles placed automatically, with AI layout matching that reads the customer's design request
- Automatic flags for long names, special requests, punctuation, and risky layouts
- Full layer-based canvas editor in the browser, so standard tags never open Illustrator
- Guided one-at-a-time review, grouped by plate color in print order
- A keyboard-driven flow that lets you approve, skip, and edit without touching the mouse
- One-click production sheets and shipping pages, golden-snapshot tested rendering core
Impact
Results
- Per Designer, Per Day
- 4 hrs → 30 min
- Orders Processed
- 5,285+
- Running Volume
- ~300/day
- Saved Per Month
- ~240 hrs
Running daily at Gem Awards. This system effectively offset a departing $32K a year role, turned up to 4 hours a day of per-designer transcription into a 20 to 30 minute review pass, and is cutting the Adobe bill because standard tags never need Illustrator. It now runs 24/7 on its own server, and the team I trained runs it day to day. I build systems the business owns.
Start with a free AI Audit.
A free 30-minute audit of the work your team repeats every day. If there is a fit, you get a fixed proposal within a week. I do not pitch you and I do not chase you.
Or email me at treyson@treysonwelch.com