LLMs ARE OVERRATED. FACTORIES NEED SLMs INSTEAD!
When people hear “AI in manufacturing,” they think of massive LLMs, complicated deployments, and costs that only Big Tech can afford.
But let’s start closer with something familiar. Take defect detection in your test and QA process.
Every factory already collects huge amounts of data: Defect logs, yield reports, FA images.
Collected. Collated. Categorised.
The problem isn’t data. The problem is:
🌀 Too much of it (really hard to analyse to extract trends & patterns)
🌀 All static (we usually look only after issues happen)
This is where AI comes in. SLMs, not LLMs.
With simple data science: data analysis, data engineering, data visualisation.
In other words, focused machine learning (ML) and developing models that:
✅ Spot anomalies in your test data before they happen
✅ Flag subtle trends and shifts in distributions that SPC charts might miss
✅ Correlate defect patterns across machines, lots, or even suppliers
None of this requires massive hardware – or a billion dollar AI system.
Most factories already have 80% of what’s needed:
👍 Historical test data (structured + unstructured)
👌 IT systems that store all this data
👨💻 Engineers who know the failure modes inside out
What’s missing is probably just an additional layer:
👨🔬 Data analysts/engineers who can clean and prepare the data
🎉 Simple ML models trained to recognize patterns faster than the human eye Don’t know where to start?
Here’s a high-level recipe:
1️⃣ PICK ONE AREA Start small. Defect detection is perfect because data is already available.
2️⃣ GATHER AND CLEAN THE DATA Consolidate your test logs, FA databases, and images. Focus on consistency.
3️⃣ APPLY THE RIGHT TOOL Use lightweight ML models (suggest open sources) to flag unusual patterns.
4️⃣ VALIDATE WITH EXPERTS The engineers need to confirm whether the AI’s flagged defects are real.
5️⃣ DEPLOY IN PARALLEL Run the model alongside your existing QA processes. Build confidence first before replacing anything.
6️⃣ SCALE GRADUALLY Once it works for one test stage, extend to others. Or move on to more sophisticated models – like predictive maintenance, process drift detection, or optimizing test times.
~~
The key point is: you don’t need to “import” AI into the factory. You already have it in the data. These models are just tools to unlock what’s been sitting around for years.
AI in production isn’t about replacing people.
It’s about giving engineers and managers sharper tools to anticipate problems, react faster, and prevent them before they happen.
The question isn’t whether you can afford AI.
The question is whether you can afford NOT to have AI on your shop floor.
~~
Need help getting started? Or more pointers? Get in touch.
Start small, validate fast, and you’ll wonder why you didn’t begin sooner.
