APEX

Industrial AI assistant for a resin-coated paper manufacturing plant.

APEX connects directly to the factory's live database and lets engineers and plant managers ask plain-English questions about production — reel output, work orders, machine performance, gas consumption, dryer temperatures, and MES status — and get back data-grounded, context-aware answers in real time.

---

What It Does

Instead of navigating dashboards and running SQL queries, operators ask APEX:

"Which machine had the most failed MES entries this week?" "Compare gas consumption between IM6 and IM7 for work order WO-TEST-042." "Show me all reels where zone 8 dropped below 185°C."

APEX pulls the relevant data from the database, injects it as context into a local LLM, and returns an answer that interprets the numbers in production context — not just repeating them.

---

Architecture

IM6 / IM7 machines
       │  (MQTT)
  Node-RED broker
       │
  PostgreSQL (reel_data table)
       │
  APEX Backend (FastAPI)
  ├── Query router   — detects intent from user message, fetches relevant DB data
  ├── System prompt  — gives the LLM full plant context (machines, reel ID format, KPIs)
  └── Ollama (local LLM, e.g. gemma3) — generates the final response
       │
  React Frontend — dark industrial chat UI

No data leaves the plant. The LLM runs locally via Ollama — no cloud API calls, no external data transmission.

---

Data Model

Every reel produced on IM6 or IM7 generates a record in the reel_data table:

Field	Description
reel_number	Unique ID — format YY<M>XXXXXXX (e.g. 256000001 = 2025, IM6, seq 1)
work_order_no	Links the reel to a production order
machine	IM6 or IM7
roller	A or B (flying splice unwind position)
reel_length	Finished reel length (m)
web_speed_average/max/min	Web speed across the reel run (m/min)
web_tension_average	Web tension (N)
web_temp_avg	Average web surface temperature (°C)
gas_consumed	Gas used per reel (m³) — efficiency indicator
zone1_average … zone8_average	Dryer oven zone temps (°C), typical 180–200°C
mes_status	SUCCESS or failure code — MES integration result

---

Query Routing

The backend detects intent from the user's message and fetches only the relevant data — avoiding unnecessary DB load and keeping the LLM context focused:

Trigger keywords	Data fetched
reel number (e.g. 256000001)	Single reel detail
WO-xxx / work order number	All reels for that work order
production, output, reel	Recent reels + production summary
speed, tension	Speed/tension analysis
gas, consumption, energy	Gas consumption report
fail, error, MES	Failed/errored reel records
Time words (today, week, month)	Adjusts lookback window automatically

---

Tech Stack

Layer	Technology
Backend	FastAPI (Python)
LLM	Ollama — runs locally (tested with gemma3)
Database	PostgreSQL + SQLAlchemy
Data pipeline	MQTT → Node-RED → PostgreSQL
Frontend	React
Data processing	pandas

---

Setup

Prerequisites

• Python 3.10+
• Node.js 18+
• PostgreSQL with the reel_data table populated
• Ollama installed and running locally

1. Clone and configure

git clone https://github.com/your-username/apex.git
cd apex

cp .env.example .env
# Edit .env with your PostgreSQL credentials and Ollama model name

2. Backend

cd backend
pip install fastapi uvicorn sqlalchemy psycopg2-binary pandas python-dotenv ollama
uvicorn main:app --reload --port 8000

3. Pull your LLM

ollama pull gemma3

4. Frontend

cd frontend/apex-ui
npm install
npm start

Open http://localhost:3000 — the chat UI connects to the backend at http://localhost:8000.

---

Example Questions

What was total production output this week?
Compare IM6 and IM7 reel lengths for the last 30 days.
Show me work order WO-TEST-001.
Which reels had MES failures yesterday?
What's the gas consumption trend over the last month?
Give me the detail for reel 256000042.