Full Transcript
00:00
Hello, this is Jim. In this video, I’m going to talk about human in the loop. Human in the loop is typically an intervention in an AI process where a human can either validate, make some acceptance or decline, or really review something before it proceeds. We see
00:15
this early on in the AI journey where you’re reviewing what came out of AI, you’re parsing it, and then sending on to another process. We also see this further down a journey where people are using AI for things maybe like reviewing a PDF and you’re extracting information
00:31
from a PDF looking for an invoice. Maybe that invoice has incorrect numbers on it or it didn’t get transformed correctly. A human can then take a look at it say yes I approve it or modify it and send it along. This particular demo I’ll be showcasing a chat using an agent. It’s
00:48
an AI agent that’s going to do research. So, whatever question we ask the chatbot, our agent is going to be looking and doing research. For example, they’re a skilled researcher, an assistant. They’re going to search the web and find reliable sources and gather
01:02
information. And then we’ll be handing that off to an a writer. It’s an agent that’s going to do writing. So, it’s specifically set out you’re a master writer. your roles that take the input from the research agent and create a well ststructured format and bring back a
01:17
summary that’s clear and very organized. I know this is a very simple structure but it’s going to help us demonstrate human and loop. I’m going to run this. We’re running on self-hosted NADN. Um it’s using um our actual NAD docker component that we have out there using
01:35
Postgress, Reddus and Caddy. So this demo is self-hosted but we are using um open AAI chat agent and so I’ll walk this through. So let’s go ahead and ask a question that we want to have researched and then summarized in writing. The question we’ll ask is what
01:52
are the latest trends in AI powered healthcare solutions? Which companies are leading the space? So you can see my workflow. The agent’s going back and forth. It’s searching the web. It’s hitting the LLM. Um, you can see here quickly that it did some research on our
02:07
behalf and it’s going to come back with some nice research. It will then be sending that over to our agent. Now, this next agent, the writer agent, that’s hooked up to a memory, but it’s also using the LLM. It’s going to summarize the information it received
02:22
and it’s going to bring it back to us. You can see it’s pretty well formulated. Let’s um make this a little larger. We used about 5,000 tokens. says AI powered healthcare solutions latest trends and it talks about accelerating adoption clinical impact of it. Here are some
02:38
leading companies. It’s talking about the healthcare pioneers in the space. The key takeaways and it has some sources. It was a pretty good output. Now let’s say we added human in the loop. And the idea here could be maybe we’re a research company. We would want
02:52
to take a look at this before we sent it to the writer. You could hit the button that says decline, send it back to the researcher and give it some notes. But in our case, we’re just going to have the option to approve it or decline it. And to do that, I’m going to actually
03:06
use telegram. So let’s take a look at what some of the options are in NAN. I clicked on to add a node. I’m going to select human and loop. And you say we have a few different systems here to add or review for response. The first one’s Discord. The next is Gmail. You can see
03:22
Google chat, Outlook, Teams. You can respond directly inside of the chat. You can send it to an email where someone has to click on a button from there. Slack and we’re going to use Telegram. Now, I’ve already configured this. I’m just using a Telegram bot that I quickly
03:35
created. I’ll go ahead and bring that in here. So, here’s my Telegram. I’m going to enable this. And then I’m basically going to put it right between our two different AI agents. Now, I have Telegram in the middle. Now that I made this change, let’s go ahead and ask a
03:51
different question. Let me copy it from another window. We’re going to ask summarize the main arguments for and against the 4-day work week, including any recent or pilot results. Same process is going to happen here. It’s going to run through and do its re. It’s
04:07
continuing to do its research. It’s going to go through and pull information and then it stops right here in the middle. It says, you can see it sent itself note. So, here in my telegram, let’s see if I can take a look at this. Um, there we go. It says, “Do you
04:20
approve this info before we send it to the writer?” And it leaves the research it found. It says, “The main argument for the 4-day workg great has some categories. It lists the sources, told me where it came from, and I have this approve button here at the end.” When I
04:33
click this approve button, it’s going to take me to a page. In this case, it’s going to then continue sending on to the writer. Now, we could have sent a message, hey, review this. Turn back in our timely response. We could have put a decline. There’s many things we could
04:47
have done. kind of adjusted it and then sent it. But in this case, it was just doing notification. And then you can see again the output that came along here, the summary arguments for a 4-day work week. Just a quick demo to showcase how the power of using human in loop can be
05:03
used. It’s a quick recap. Human and loop is a great way to get some feedback in the middle. It’s really nice to add some gate between something or think of it as speed bump as you’re doing work with AI. This was all done using NAD. If you’ve got more question, feel free to reach out.
Let's work together.
Partner with Augusto to streamline your digital operations, improve scalability, and enhance user experience. Whether you're facing infrastructure challenges or looking to elevate your digital strategy, our team is ready to help.
Schedule a Consult