This episode of the Pipeliners Podcast features a conversation with Samer El Issa of FloSeer about applying AI to solve real-world challenges in pipeline and measurement operations. The discussion covers how different types of AI, including large language models, mathematical models, and agent-based systems, can be used together to deliver practical and reliable outcomes, as well as the key considerations involved in building effective AI-driven applications.
Listeners will gain insight into how AI is evolving beyond basic use cases, along with the importance of context, data control, and system design in turning AI into a tool that drives measurable operational value.
Creating a Value-Delivering AI Application Show Notes, Links, and Insider Terms
- Samer El Issa is the Managing Director of FloSeer, with over two decades of experience in oil and gas automation, specializing in flow measurement, custody transfer, and the application of IIoT and AI to improve operational accuracy and efficiency. Connect with Samer on LinkedIn.
- FloSeer is a software company that provides real-time data analysis to validate custody transfer volumes, detect discrepancies, and improve accuracy and accountability in liquid transfer operations.
- POEMS (Pipeline Operations Excellence Management System) is a software platform designed to support pipeline operators in managing safety, compliance, control room operations, and field activities.
- AI (Artificial Intelligence) is the broad field of using algorithms and computational models to perform tasks that typically require human intelligence, such as pattern recognition, language processing, and decision-making.
- LLM (Large Language Model) is a type of AI that processes and generates human language using probabilistic methods based on patterns learned from large datasets.
- Generative AI is AI capable of creating new content (text, code, etc.) by predicting and generating outputs based on learned patterns from data.
- Machine Learning (ML) is a subset of AI focused on building models that learn patterns from data to make predictions or decisions without explicit programming.
- Neural Networks are computational models inspired by the human brain that are used in machine learning to identify patterns and relationships in data.
- Deep Learning is a subset of machine learning that uses multi-layered neural networks to analyze complex data such as images, speech, or time-series data.
- Transformers (AI Architecture) are a type of neural network architecture that enables efficient processing of language by understanding relationships between words through attention mechanisms.
- Tokenization is the process of breaking down language into smaller units (tokens) such as words or phrases so AI models can process and analyze text mathematically.
- Tokens are the individual pieces of language (words, subwords, or phrases) that AI models use as the basic units for processing and prediction.
- Probabilistic Model is a system that predicts outcomes based on likelihoods and statistical probabilities rather than deterministic rules.
- Deterministic Model is a system that produces consistent, repeatable outputs for a given input, commonly used in mathematical or rule-based AI applications.
- Pattern Recognition is the process of identifying trends, correlations, or anomalies in data, often used in analyzing operational data such as pressure or flow trends.
- IIoT (Industrial Internet of Things) is the integration of sensors, devices, and systems in industrial environments to collect and exchange data for monitoring and optimization.
- Industrial Automation is the use of control systems and technologies to operate industrial processes with minimal human intervention.
- Control System Engineering is the design and implementation of systems that monitor and control industrial operations such as pipelines, flow measurement, and terminals.
- Flow Measurement is the quantification of liquids or gases moving through pipelines, critical for operations, custody transfer, and compliance.
- Custody Transfer is the process of measuring and transferring ownership of oil or gas, where measurement accuracy directly impacts financial transactions.
- ATG (Automatic Tank Gauging) is a system that uses sensors (often radar-based) to measure liquid levels in storage tanks for inventory tracking.
- Material Balance is the process of accounting for all inputs, outputs, and storage within a system to detect discrepancies such as losses or measurement errors.
- Real-Time Data Processing is the continuous analysis of data as it is generated, enabling immediate insights and operational decisions.
- Accuracy (Measurement Accuracy) refers to how close a measured value is to the true value, critical in custody transfer and reconciliation processes.
- Drift (Measurement Drift) is the gradual deviation of measurement accuracy over time, which can lead to discrepancies between systems.
- AI Agent (Agentic AI) is an AI system that can autonomously perform tasks by breaking down problems, executing workflows, and interacting with other tools or agents.
- Agentic AI is the concept of using AI agents that can independently plan, execute, and complete multi-step tasks without constant human input.
- Orchestrator Agent is an AI component that coordinates multiple agents, managing workflows and combining outputs into a final result.
- Workflow Automation (AI Context) is the use of AI agents to execute a sequence of tasks or processes automatically, often across multiple systems.
- Prompting is the process of giving structured instructions or inputs to an AI model to guide its output.
- Context (AI Context Setting) is the information provided to an AI model to shape how it interprets inputs and generates responses.
- Guardrails (AI Guardrails) are constraints placed on AI systems to control outputs, restrict access to data, and reduce risks such as incorrect or unsafe responses.
- Hallucination (AI Hallucination) is when an AI model generates incorrect or nonsensical outputs that appear plausible but are not grounded in actual data.
- RAG (Retrieval-Augmented Generation) is a technique where an AI model retrieves relevant data from a defined dataset to improve accuracy and reduce hallucinations.
- Fine-Tuning (AI Model Training) is the process of retraining an AI model on a specific dataset to improve performance for a particular use case.
- Local LLM (On-Premise AI Model) is an AI model deployed within a private network, ensuring data does not leave the organization’s environment.
- Open-Source Models are AI models with publicly available code that can be modified and deployed independently.
- GPU (Graphics Processing Unit) is specialized hardware that enables high-speed parallel computation, essential for training and running modern AI models.
- Air-Gapped System is a secure system isolated from external networks, often used to protect sensitive data in industrial environments.
- API (Application Programming Interface) is a mechanism that allows different software systems to communicate and exchange data.
- Leak Detection (Pipeline Operations) is the process of identifying unintended releases in pipelines using data analysis, pattern recognition, or monitoring systems.
- Spreadsheet-Based Reconciliation (Legacy Practice) refers to manual or periodic data reconciliation (e.g., in Excel), often slower and less reliable than automated systems.
- Bias (AI Bias) is the tendency of AI models to reflect the assumptions, data, or perspectives of their training sources or developers.
- NotebookLM is a tool that allows users to constrain AI responses to a specific set of uploaded documents for more focused and accurate outputs.
- Anthropic is an AI company that develops large language models such as Claude.
- Claude (Anthropic Model) is an example of a commercial large language model known for strong capabilities in coding and agent-based workflows.
Creating a Value-Delivering AI Application Full Episode Transcript
Welcome to the Pipeliners Podcast, Episode 433, sponsored by EnerSys Corporation, providers of POEMS, the Pipeline Operations Excellence Management System—operations and compliance software for the pipeline operator to address safety, program management, control room management, and field operations.
Find out more about POEMS at enersyscorp.com.
The Pipeliners Podcast, where professionals, bubba geeks, and industry insiders share their knowledge and experience about technology, projects, and pipeline operations.
And now your host, Russel Treat.
Russel Treat
Thanks for listening to the Pipeliners Podcast. I appreciate you taking the time.
And to show that appreciation, we give away a customized YETI tumbler to one listener every episode. This week, our winner is Randy Smith with New Jersey Natural Gas. Congratulations, Randy. Your YETI is on its way.
To learn how you can win this prize, stick around until the end of the episode.
This week, we’ll speak with Samer El Issa of FloSeer about using AI to create applications that deliver real operating value.
Howdy, Samer. Welcome to the Pipeliners Podcast.
Samer El Issa
Hi, Russel. How are you?
Russel Treat
I’m doing great. It’s a wonderful Monday.
Samer El Issa
Absolutely. We’re getting our last little bit of cool weather before the summer starts to set in, so I’m enjoying it.
Russel Treat
Yeah, it was a little cool this morning.
Well, look, let me start like I normally do. I’m going to ask you to do a little bit of an introduction. Could you tell us a little bit about who you are and what you do, and kind of how you got into that job?
Samer El Issa
Yeah. So my name is Samer El Issa. I’m the Managing Director of FloSeer.
My background is in engineering, and I started my career off doing industrial automation. I worked at GE and then Emerson doing control system engineering and solutions. Then I did IIoT and data management, started getting into some management work, did my MBA, and then I got the opportunity to move to Houston in a global leadership role with Emerson.
Before that, this was all in Canada, in Edmonton and Calgary. Then with Emerson, I was within the flow measurement division, managing the control system piece and the IoT solutions for flow measurement. Then I decided to leave Emerson and start my own business and go at it on my own.
Russel Treat
Tell us a little bit about that business. What is that, and what do you guys do?
Samer El Issa
Yeah. So the business, FloSeer, is a software business, a software and AI business, within the measurement space.
There are three partners, and we built it all around a specific customer need. It was a big NOC that had a specific issue, and we found that issue exists across the globe, so we built the product around that. It’s been running with millions of barrels going through it a day, so it’s hardened and ready to go.
The issue we’re trying to resolve is that most operators use custody transfer meters for custody transfer and use ATGs, basically, with radars and all that equipment, but they just use it for inventory. Although API allows you to use level changes for custody transfer, most operators don’t use it that way.
From what I found and what I’ve observed, they basically reconcile between the two. They do a material balance once a month, maybe once every two weeks, at most once a week. What this software does is perform that material balance in real time.
So it’s tank by tank, valve by valve. Every time a valve opens or a tank level changes, it tracks it end to end and compares it in real time. Usually, with custody transfer, you get 0.3% accuracy. Here, we usually get 0.02% to 0.03% comparison between those two systems.
So we can detect any drift between them. You have a backup system. If there is a dispute, then you basically have two independent measurements and can compare the two. So there’s a lot of good value and features there, and you get rid of doing it in Excel.
Russel Treat
Yeah, yeah. Doing it in real time, using AI to look for root causes versus doing it in a spreadsheet when you get around to it.
Samer El Issa
Absolutely.
Russel Treat
It’s interesting, right?
Well, look, I want to pivot a little bit and talk a little bit about AI in general. I want to build a little context for our conversation.
I’ve been doing some work in AI, and I’ve been reading some books and trying to understand what it is, how to use it, what its limitations are, what its value propositions are, and all of that. Then I ran into you. I think I ran into you at Corpus Christi Measurement School.
Samer El Issa
That’s correct.
Russel Treat
And you were showing me your application. I don’t know, it took me a minute to really understand what you were doing and think, “Oh, there’s the value.” It took me a minute to get there.
But I think what’s interesting about what you’re doing, and why I wanted you to come on, is that you’re actually using AI to solve a very specific technical and numerical problem, which is different from what a lot of people are thinking about when they’re thinking about AI. They’re thinking about LLMs and policies and procedures and instructions and things like that, not necessarily the mathematical pieces.
So anyway, as a tee-up, what I want to talk to you about is: how do you go about understanding AI well enough to build something that actually creates value?
Samer El Issa
Yeah. So I’ll admit, the first time I used AI, I basically thought it was a glorified Google search. Embarrassing now, but that was my first thought.
I realize a lot of people now basically use it that way. But I continued experimenting and playing around and trying different things and watching YouTube videos. There’s a lot of material out there, but it’s all over the place. And if you keep up with AI, there are changes happening basically seven times a day. So it’s very, very fast-moving.
I spent a lot of time essentially trying to understand how it works. Eventually, I got a laptop with a GPU, a really powerful GPU, and I ran an open-source model on it. I started building an application on that local model, so it was completely air-gapped, and I could do whatever I wanted with it.
That’s really how I got to experiment and learn.
Russel Treat
Yeah. For people who are hearing this, I think your comment about AI being a glorified Google—or maybe Google on steroids—is something a lot of people can relate to. And certainly, there’s a lot of value in using AI that way, but that’s really just taking the crust off the biscuit of what you can actually do with it.
So maybe we ought to break it down a little bit and talk about the different types of AI. Everybody’s hearing about LLMs, or large language models, but that’s really only one class of AI, right?
Samer El Issa
Yeah, that’s correct.
AI has been around for a while. AI is not new. We had machine learning, neural networks, deep learning, and all of their applications. I was doing numerical methods and neural network programming in Fortran. So it’s been around for a while.
Russel Treat
In the seventies.
Samer El Issa
Exactly. And all of that has been around. There’s computer vision, there’s speech recognition, and all of that. But what’s new, and what came out that basically transformed all of this, is really two things: something called transformers, and the ability to take language and transform it into tokens, along with powerful devices, which are GPUs, the graphical processing units.
Those are really built for gamers, but you can utilize them for this purpose. These things allowed for generative AI. So what it can do is read language and then generate results.
Traditional AI might look at an email and tell you whether it’s a threat or not based on pattern recognition. This can write the entire email for you.
Russel Treat
Well, and you said a couple of things I think are kind of important to understand, and that is the idea of transformation of language.
Basically, you take a language like English and turn it into tokens. A token could be a full word. It could be a fractional word. It could be a couple of words together. But it’s breaking language down into the essential parts—something that’s big enough to have meaning, but small enough to have a discrete meaning. Right?
Samer El Issa
Yes.
Russel Treat
And that’s not as easy as it sounds, particularly in English, where we use the same word to mean different things. Which is why you have to tokenize it, because you’ve got to build context around it.
If I use the word “foot,” do I mean a unit of measure, or do I mean the thing at the end of my leg? So you have to build context. “Leg” and “foot” could go together as a token, and “inch” and “foot” could go together as a token, and they mean different things.
That’s kind of a huge oversimplification, but that tokenization allows you to start using math against the language.
Samer El Issa
Absolutely. And the math you’re using is statistics. You’re saying when you see these two things together, it most likely means this other thing.
Russel Treat
Yes, and that’s how it works. It’s pattern analysis. So when you say, “I am super,” it then looks probabilistically at what the next word might be. Is it “Superman”? Is it “super excited”? Right?
Samer El Issa
Yeah, exactly.
Russel Treat
Or “superintendent,” right?
Samer El Issa
Exactly. So basically, the AI, the LLM, doesn’t understand. It just probabilistically tells you what the next word is. And that’s what this is all built around.
Russel Treat
Yeah. And I think that’s one of the things that bothers me when people talk about AI developing intelligence or sentience and all that kind of stuff. That bothers me because I think that’s marketing hype.
It’s just math. It’s doing a whole lot of math, and it’s doing it very, very fast, but it’s still just math. It’s just an algorithm. So like any other algorithm, you put something into it, you’re going to get something out of it, and it’s just a process.
Samer El Issa
I completely agree. I think the danger is that people rely on it thinking that it does understand, and it doesn’t.
Understanding how the output could be wrong is also very important, because now you need to control the output so it doesn’t cause you harm, and you need to put the right guardrails in place for the AI to be able to use the system without it hallucinating.
Russel Treat
Yeah. And all that means is that I feed information into the algorithm, the algorithm gives me an answer that makes no sense.
Now, the interesting thing about that is mathematically, it does make sense, right? Because you gave it data, it processed the data, and it gave you a result. It just wasn’t the result you were looking for.
So you’ve got to go back into the black box and tweak the black box to give you the answer you actually wanted.
And that process with LLMs has two pieces to it. One is, you’ve got to build context for the LLM and say, “I’m talking about units of measure,” and then that’s going to cause it to interpret the token “foot” differently. Or, “I’m talking about body parts.” That’s called building context.
Then you have to give it guardrails. Context is, “It is this,” and guardrails are, “It is not this.”
Samer El Issa
Yes, absolutely. And there are guardrails also outside the system.
If you connect the system into a database, for example, you could put in a guardrail that it can’t access certain information, or it can only read certain things.
So the guardrails could be inside the AI, the LLM itself, or outside. And don’t rely solely on the guardrails within the LLM, because with a big enough context, it could forget and basically ignore them.
Russel Treat
Yeah. So I want to ask a little bit of a different question.
Given your background with Emerson, I’m sure you’ve got a lot of experience with this. Typically, when we talk about AI in the current day and age, we’re talking about an LLM. But there are other kinds of AI that are purely math—doing statistical analysis, looking for patterns, all that kind of stuff.
What are the key differences between an LLM and a mathematical type of AI?
Samer El Issa
So the mathematical type of AI—the one that already existed, like deep learning and neural networks—is better suited for math and for pattern recognition.
LLMs should not be used for that. A lot of people make that mistake. If you ask an LLM to do a mathematical calculation for you, you may be disappointed. It might be the right answer, but I wouldn’t count on it.
So when we built things for flow, within the agent, it goes through the pieces it needs to do. If it’s an actual calculation, then just do the calculation. I don’t need the LLM for that. It’s a tool. Pass it over there and get the calculation.
If it’s pattern recognition, it can go to a machine learning algorithm. Now, if I want to analyze a report, summarize a report, read it, or pull out relevant pieces from it, an LLM is excellent for that.
If I want it to think through and troubleshoot something in a more human-like way, an LLM is great at that. But then the technical pieces can get handed off to the mathematical pieces.
Russel Treat
I just want to decode a couple of things for people that might not have the language.
When you’re talking about pattern recognition, basically what you’re talking about is: I’ve got a string of numbers—say, pressure over a long period of time—and I’m looking at how that pressure changes.
Pattern recognition is calculating distributions, means, modes, and looking at what happens when valves are in one position versus another. It’s looking at those things mathematically.
That kind of AI tends to be deterministic, meaning it gives you a specific answer, versus an LLM, which is probabilistic because language is probabilistic.
I have a thought, but I don’t necessarily say exactly what I think. That’s probability, right? So LLMs are probabilistic, and therein lies the difference.
The other term you used was “agent.” So I want to talk to you a little bit about what an AI agent is and why that matters in this conversation.
Samer El Issa
Yeah. So this is where some of the other use cases for AI really pop up and where it starts to add value.
Agentic AI, or an agent, is basically an LLM with other functions that can execute an entire operation on its own. And it could involve multiple agents.
So the agent decides what it wants to do. You ask it to do something, it breaks it down, and it has a workflow that it walks through.
As an example, let’s say we have a material balance issue. It passes that on to the next agent. That agent now does a leak detection search, or it looks at the data and detects that there’s a mismatch somewhere, and then it decides where it wants to go next.
So the LLM can now perform a full function end to end and give you the result. It’s not just the typical chat interaction where you ask the LLM something and it gives you a response. It could be running in the background twenty-four hours a day, looking at results and doing the functions it needs to do.
Russel Treat
Right, right. So another example of this would be in the software world.
If I were going to use an LLM to write code and build an application, I could have one agent that is the designer. I could say, “I want you to design an application that finds relevant Wikipedia articles and summarizes them so I understand the problems I’m trying to solve.”
Then I could have another agent that, given that problem statement, builds a piece of software that does that. Then I could have another one that says, given this problem and this software, build me a set of tests to test it.
So you’d have a software design agent, a software programming agent, and a software test agent, all working together as a team. Is that the right way to think about it?
Samer El Issa
Absolutely. And a lot of the best AI programming LLMs out there—Claude is probably number one right now—actually have an agent built into them.
Anthropic, the company that builds Claude, has strong models available. So when you say, “Build me an application that reads my PDFs,” it actually goes through and creates multiple agents, and each agent goes out and performs a function.
One agent works on the front end. Another agent works on the back end. Another agent tests it. They all work together, report the results, and there’s an orchestrator agent that pulls everything together and gives you the end result.
Russel Treat
So if you think about it as a team, each agent is a person on the team. If you think about it as a factory, each agent is a machine in the factory that takes a certain input, performs a process, and generates an output.
Samer El Issa
That’s correct.
And agentic AI can be open-ended, where you give it a lot of control to decide on its own, or you can make it much more deterministic.
There’s no such thing as fully deterministic in this space, but you can say, “You do this, and when you’re done, hand it off to that agent, and that agent does that.”
Usually, for oil and gas, you’d want it to be much more deterministic than, say, an agent that builds a consumer application.
Russel Treat
Yeah, that’s right. Exactly.
So let’s pivot a little bit. We’ve talked about LLMs and mathematical AI. We’ve talked about agents.
Now let’s talk about building an application. We’ve talked about context and prompts and training—well, we haven’t really talked about prompts and training.
Sorry, I’m babbling a bit, because one of the things I find about AI is that it’s very hard to get your semantics straight and grounded, because it’s changing so quickly. What works as a mental model today might be completely out the window in six weeks.
So what do you need to be thinking through when you start thinking about building an application?
Samer El Issa
Yeah. So there are many things involved there.
You mentioned prompting, and context comes into prompting. If you say, “Make me a meal,” just like if you go to a restaurant and say, “Make me a meal,” you’ll most likely be disappointed.
You need to describe what you want in that meal. What do you like? What’s the setting? Are you on a work lunch, or are you taking your daughter out for dinner? The context is important.
You tell it exactly what you want: “I like this, I don’t like that, do this.” Usually positive reinforcement works best.
And there are many techniques. There’s one-shot prompting or few-shot prompting, where you give it an example and say, “Here’s how this is done. Do it this way.”
There’s chain-of-thought prompting, which is how you break down the task. You say, “Do this, then do that, then after that…” and you walk it through.
There are many techniques. The best way to figure this out is just to try.
Russel Treat
But as you said earlier, you’ve got to play with it, right?
Samer El Issa
Yes.
Russel Treat
And what works for Claude may not work with Gemini, right? Each one is a little different. And what works with ChatGPT 4.0 might not work with ChatGPT 5.0.
Samer El Issa
Yes, which is a whole other kind of problem.
That’s a big problem, because now you’ve spent all this time building this amazing agentic AI, and then the company you’re using decides to shut down the old model or make a change overnight, and suddenly your whole system falls apart.
So it takes a lot of work to keep that maintained.
Russel Treat
Yeah, exactly.
So a lot of people will say, “I don’t want an AI looking at my data,” right? Because I’m basically going to give my data to the world if I give my data to an AI.
How do you address that concern?
Samer El Issa
Yes, and that ties to the previous point about the model changing.
The way we’ve addressed it, and what I believe is going to be the future, is that a lot of this is going to be done with local LLMs.
Every AI company you connect to has a checkbox somewhere that says, “Do not use my data to train the model.” Do I know for sure that that’s what’s going to happen? Not really.
I’ve heard of examples where people gave an API key and then it made it out somewhere. So I don’t personally trust that.
What we’ve done within FloSeer is use a local LLM with open source. Up until recently, the open-source ones were often called small language models. They were smaller. They weren’t as powerful as the ones running on million-dollar data centers, but for a specific task, they do really, really well.
And it’s yours. Your data doesn’t leave the network. Nothing changes until you want to make that change.
About a week ago, they released a model that is basically almost as good as the cloud-based ones. So that race is happening as we speak, and no one knows where it will go.
Russel Treat
Yeah, yeah, yeah. I think that’s absolutely right.
I think the other thing we talked about was hallucination. One of the things I’ve been reading a fair amount about recently is how do you verify the results of a model, and in particular, how do you automate that verification?
That’s a fascinating topic in itself. Some of the conversations are about what you do manually versus using an AI to verify an AI. That’s fascinating, particularly because if you want to do a very large set of tests, you have to automate that process.
But if you build a very large set of tests into an LLM, there are all kinds of things that can go sideways where you just kind of regenerate a bad answer.
So it’s a very fascinating concept. Have you played with that at all yet?
Samer El Issa
Yeah. So first of all, before that, there’s something that helps with this, which is a technique called RAG.
You can retrain a model or fine-tune it, but there’s also retrieval-augmented generation. What that does is connect the model to data that you give it. It’s kind of like an extended memory.
So it does not hallucinate in the same way, because it goes right to that source of data to get the information. That eliminates some of that problem.
Now where it gets tricky—and that’s something I’ve been playing with—is using AI to build a test model before building the actual solution.
And you’re right, there is a risk that the same mistake gets made in the testing protocol that gets made in the actual application.
So the only way I’ve found so far to address it—and this is moving very quickly, so things will change—is to test it in real life with another manual method of testing and verifying it that way.
Russel Treat
Well, with the mathematical side, it gets easier because you can kind of build a set of inputs and outputs and say, “This is my reference result.”
With the language models, that’s more difficult, but you can do the same thing.
I think what’s fascinating about this whole subject is that large language models, which have been trained against everything on the internet, tend to have the biases of their programmers kind of incorporated into them.
The person building and teaching the LLM how to ingest data and respond to prompts is giving it criteria. It would be impossible for me not to inject my biases into that training.
Samer El Issa
Yes, that’s a big risk out there, and it does exist.
If you go ask a Chinese LLM a question and then ask an Anthropic model—a U.S.-based LLM—the same question, you may get different results based on the biases that are inherent in their culture.
In these large language models, the culture of the organization creating it tends to get embedded into the AI. Not deliberately, but just as a function of how the system works.
Russel Treat
Yes. And the guardrails they’ve set in place are different.
Samer El Issa
Exactly.
Russel Treat
Because it’s based on their belief system, right? And what they think is correct.
I think what’s interesting, though—and I’d like to hear your take on this—is my thesis that the more I can work with a smaller set of data, the more I can dial in the answer and make it technically correct.
So, for example, if I were just worried about operational safety, I could focus on a known body of work—say, U.S. OSHA requirements and a certain set of books I wanted to feed the AI—and then out of that, I’d get a much more specific response because the universe of supporting information is smaller.
Samer El Issa
Yes. And by the way, Google built a tool specifically for this. I don’t know if you’ve used it—NotebookLM.
What it does is let you upload a specific set of documents, and then it references absolutely nothing outside of those documents. That limits it and gives you a much more focused answer.
So you’re absolutely right. That’s the way to do it.
And the same thing applies if you do it with a local model. The more you constrain it, the better results you have.
If you give it a lot of data and a lot of context, there’s a limit. It’s like a human. If I can remember this conversation, it’s because it’s all within my mind right now. But if you suddenly read one hundred numbers to me, it would overwhelm my system and I wouldn’t be able to keep up.
If we have a seven-hour conversation, I’ll start forgetting where we are. It’s the same thing.
Russel Treat
Yeah, exactly, exactly. There’s an optimal amount of context, and then there’s a suboptimal amount, either too little or too much.
Samer El Issa
Exactly.
And then you also have memory, which the LLM has in its own way. If you want your Social Security number, for example, it’s somewhere in the back of your mind. You have to go back and retrieve it.
So if you contain within the LLM the right amount of information, the right context, and you control all of that with the right prompts, that’s where you get the results you actually want.
Russel Treat
Yeah. So I think the takeaway from this is: if you’re trying to build an application to deliver value and you’re going to leverage an LLM, context and control are critical.
And in a lot of cases, for those of us doing technical things—whether that’s regulatory compliance or analytics of math—less information that I’m training on is better.
Samer El Issa
Yes. I mean, you’ve got to have enough to get a reasonable response, but to the extent you can control it, the better.
If you want to train it, then yes, you want the maximum amount of consistent data to train on. But once it’s trained and ready, you want to give it the right amount of information to get the right output.
Russel Treat
Yeah. Well, I think that’s a whole other conversation, Samer, which is the training they do to build an LLM versus the training you do to build an application. Those are different things.
Samer El Issa
Yes. And in most cases, you don’t need to retrain or fine-tune the LLM.
It’s a good idea if, let’s say, you want to build a customer service application and you’re Google or Amazon and you have a lot of data. Then you can retrain it based on your data set, which is a huge data set, and it becomes really good at that.
But if you ask it about the weather, it has no clue. It will get lost. So it becomes an expert in something very specific.
That’s another way to do this.
Russel Treat
Yeah, and that actually gets to where the real opportunity is in this whole AI space. It’s not in these big models. It’s in the applications that deliver real value, that make us more effective as engineers.
It just becomes another tool in the belt.
I remember very well in high school—this was in the mid-seventies—when the math teachers were like, “No, no, no, you can’t use a calculator. You have to do it all by hand because we want you to learn it.”
Shoot, nobody does that now, right?
Times change.
Anyways, well look, kind of coming to the end of our time, this has really been a great conversation, Samer.
How would people find you if they want to quiz you about this stuff? What’s the best way to get in touch with you?
Samer El Issa
So I’m available on LinkedIn. I’ll put in the link. And on FloSeer.com, you can find a link there as well.
Russel Treat
Yeah, that’s FloSeer, F-L-O-S-E-E-R dot com.
Samer El Issa
That’s correct.
Russel Treat
Right.
All right. Well look, as usual, we’ll put all this stuff into the show notes. So if you want to find Samer, just go to the Pipeline Podcast Network website and find this episode. You can find Samer and his contact information there, and we’ll put some links in the show notes so you can learn a little bit more about his technology and what he’s doing for liquid terminal operators to help them control their measurement.
Hey, thanks so much for coming on.
Samer El Issa
Thank you very much. It was a great conversation.
Russel Treat
I hope you enjoyed this week’s episode of the Pipeliners Podcast and our conversation with Samer.
Just a reminder, you should register to win our customized tumbler with a blazoned Pipeliners Podcast logo right on front. Simply visit PipelinePodcastNetwork.com and enter yourself in the drawing.
If you have ideas, questions, or topics you’d be interested in, or if you’d like to be a guest, or if you know somebody who you think should be a guest, please let me know either on the Contact Us page at PipelinePodcastNetwork.com or reach out to me on LinkedIn.
Thanks for listening. I’ll talk to you next week.