Make It Possible
The Backup Failure
I once tried to explain to a tech guy why we need a simple solution, and I was shamed. He said there must be a million pieces of software out there that can fix exactly the “primitive task” I mentioned. He was right.
But another piece of the puzzle is that the everyday work of staff in the construction industry is tied up with heavy, time-sensitive work. Shared drive connections or authorization issues happen all the time, but we can’t afford the luxury of reinventing the whole system. A solution that is straightforward enough to not break, and simple enough that no learning is required, is what we need.
AI Limitation 2: The AI that needs prep talk (Continuation of Limit 1)
I was drained after yelling at the AI. So I changed the approach.
I said I would share AI limitations and use in my adaptation to hospital and equipment planning.
This one is the continuation of the inaccuracy problem and how I fixed it by giving prep talk.
In my last post I shared how I was emotionally drained after yelling at an LLM (the chatbot behind ChatGPT and Gemini) that kept giving me obsolete and inaccurate data when I asked it to compare ultrasound machines. So I changed the approach entirely.
I stopped pushing the LLM harder. The fix is doing the groundwork before I ask the question.
Instead of letting the LLM search freely, I went to the manufacturer product pages myself, copied the URL and pasted it directly. No searching. No guessing. Just read this page and tell me what is on it.
I set strict rules on the output format and asked it to keep the exact wording from the page.
But I also gave it room to work with whatever was actually there. I used prompts like “extract MEP details that is architecturally significant listed in the document” rather than forcing it to fill a fixed template.
Working with a LLM is very much like working with a smart intern. Fast. But clueless about the expectations and it can’t read your angry face, yet.
If you don’t give them the right information but still demand an answer, the only way out is they make it up or they go silent.
This is doable with LLM but it needs the same boring prep talk every time so I baked the workflow and prompt into the cutsheet generator of Equipia. 2 sequential AI: one to gather the correct source data, one to analyse and extract it.
You may try copying the website with context into the LLM as I mentioned in my last post and see the difference.
If you haven’t had a chance to read the previous post, here is the link: [LinkedIn post link]
Someone you know has been through this? Forward it to them.