Vedant Dwivedi, Co-Founder & CEO at AndAI Hub. AndAI Hub is a cutting-edge AI marketplace that connects businesses with the best AI plugins and solutions tailored to their specific needs. They cater their services to startups, MSMEs, and enterprises by understanding each client's unique business domain and recommending optimal AI integrations.

Their platform showcases a curated selection of third-party AI products and services, alongside their own in-house developed plugins, agents, workflows, and solutions. At AndAI Hub, they pride themselves on delivering customized AI solutions for every use case, ensuring their clients harness the full potential of AI to drive innovation and growth in their respective industries.

Listen to the episode on Spotify, Apple Podcast, Podcast addicts, Castbox. You can also watch this episode on YouTube.

In this episode of the AI Minds Podcast, Demetrios welcomes Vedant Dwivedi, co-founder and CEO of AndAI Hub, to explore how real-time AI is transforming enterprise operations.

Vedant reflects on his journey from leading data science at Reliance Jio in India to launching AndAI Hub in Australia, driven by the need for AI frameworks that support compliance, low latency, and real-time decision-making.

They unpack the evolution from batch to real-time AI, and why traditional models struggled to meet enterprise demands for speed, moderation, and security.

Vedant shares how AndAI Hub suite—Builder, Sigma, and Flow—helps enterprises catalog scattered data, deploy focused AI agents, and design workflows that drive automation and insight.

He illustrates this through examples like manufacturing voice agents and a recruitment system using voice, video, and code assessments—built entirely with AndAI Hub stack.

The episode also highlights how Deepgram powers voice components, enabling fast, moderated interactions in real-world AI use cases.

Listeners will gain practical insight into building scalable AI systems, managing data readiness, and creating intelligent workflows that unlock true enterprise value.

Show Notes:

00:00 Challenges in AI Model Deployment

04:55 Image Filtering: Pre or Post-Prompt?

09:16 Real-Time Voice Insights for Manufacturing

12:17 "Centralized Data Solution with Builder"

16:51 AI-Enhanced Job Recruitment Tools

17:43 Candidate Evaluation Workflow Summary

More Quotes from Vedant:

Demetrios:

Welcome back to the AI Minds Podcast. This is a podcast where we explore the companies of tomorrow being built AI first. I'm your host Demetrios and this episode, like every episode, is brought to you by Deepgram, the number one voice API on the Internet today, trusted by the world's top conversational AI leaders, enterprises and startups. Some of these folks you may have heard of like Twilio, Spotify, NASA and Citibank. In this episode I am joined by the co founder and CEO of AndAI Hub Vedant. How are you doing today?

Vedant Dwivedi:

I'm doing awesome, thanks for asking. How are you?

Demetrios:

So Vedant, you were a data scientists in India, you've since moved to Australia. What were you doing as a data scientist?

Vedant Dwivedi:

So I was working in India with Reliance jio. I was a senior data scientist. I was managing a team of around 21 people and my day in, day out job was to actually take data science application into production system. And during that time I understood that real time data science applications will be termed as artificial intelligence in future. So we are talking about 2023. This is the time when GPT also came into the picture and that time it was more around working with The NLP libraries, TensorFlow and small sets of large language models. It was not exactly the large language models which we use currently. And I understood that we need to have a framework, certain sets of framework which can actually help with the compliance, which can actually help with the latency and

Vedant Dwivedi:

That's the reason we were struggling with and that is how I thought we should have a framework. And IHUB is the real time framework which helps you with that.

Demetrios:

When you say real time, what do you mean?

Vedant Dwivedi:

So generally traditionally what happens was that in data science application lot of the training which happens is in the batch processing system you get a lot of data and that ton of data is processed in the batches. But when you have to do a real time data science application, you need to analyze the data in a real time. That means that you don't have an option to process it on the later stages. After 24 hours or 36 hours, you have to process it in the real time and you have to deliver the results in a prediction, forecasting or whatever manner in a real time. And that becomes a real time data science application when the data is streamed in and data out is a real time intelligence, not a batch processing intelligence.

Demetrios:

Okay, fascinating. And so you saw while working with large language models that there was a lot to be hoping for when it comes to real time Applications. I imagine back in those days, if it was pre chatgpt, there wasn't much as far as frameworks and even real time. Anything with large language models.

Vedant Dwivedi:

So earlier what we used to do was that we used to have encoder model, decoder model, sequence to sequence model and these were typically running on jupyter notebooks. And if you have to take anything into the production system, it was really hard to make a REST API out of it. So we say that there is a saying in data science, if you have made an API and it heads to the production systems, that is when the real test happens for an artificial intelligence. And that was really critical to take forward because you need GPU inferences in order to do that. And post GPT era it was absolutely uncommon. So most of our work was around the R and D and if possible, if we took even one model into the production, it was batch processed, not real time.

Demetrios:

Because there are so many complications that come in when you are doing it real time. Can you explain a few of those for me?

Vedant Dwivedi:

So even in real time right now there are a lot of problems. Let's say there are bag of stop words which you use in a real time AI application, you want that some certain things should be moderated and this particular kind of content AI should not be able to understand and it should not respond to some particular kind of requests. So these kind of real time decisions, you can't rely on the intelligence. You need to have a logic layer which separates the intelligence layer from the logic. Let's say if I ask the AI to create me the image, I want that image to be moderated. And this moderation takes quite a lot of time to implement on a logic layer. Because intelligence, the AI itself, the neural network, it doesn't do get the understanding of what to block and what to produce.

Vedant Dwivedi:

It is always at the top of it, you have a layer of logic which decides what to produce and what not to. So that is something which was extremely critical and that is what we implemented.

Demetrios:

And are you doing this before the prompt hits the large language model or are you doing it afterwards? Or a little bit of both, because I can imagine that I go to my favorite image generator and I want to create some kind of funny image that is not allowed and I throw in a prompt. The prompt can get automatically shot down before it even goes to the image generator. And that would save everyone a lot of money because you're not using any type of resources after it gets shot down. Or it can sneak through maybe that initial filter, the image can be made and then on the back end, once the image is made, you have another filter that says whoa, how did this make it through? And then you shoot it down and it never actually sees the user. But there are two sides of the story maybe, or two areas, surface areas that you can be filtering.

Vedant Dwivedi:

So you have to have a guardrails on both the sides, which is pre and post. So it is pre analysis. Also there has to be a guardrail which just checks whether the input prompt is not having any stop words. But even if by any chance that prompt is extremely long and it was somehow missed during the time of the pre analysis. So after the post analysis, once the content is generated, whether it is a text content, a video content or image content, you need to have a post layer as well. So it's a post guard liligard drill mechanism where you just check whether the generated content again adheres to the policies or not. So it is pre and post both the way. That's how you have to do it.

Vedant Dwivedi:

Generally all the models which we consume right now, whether it is Gemini or whether it is Llama, whether it is GPT, they do have some certain moderation on themselves. But still, when you use these API directly into enterprises, it's best to have a guardrail mechanism in the pre and post. And that is what andehive provides out of box.

Demetrios:

We've seen the stories of the car dealership that is a Chevy dealership and it recommends Teslas and says Chevys aren't as good as Teslas. Or the Air Canada incident where it's selling back a ticket for a dollar or selling someone a ticket for a dollar. And that's where you realize maybe we should not be so cavalier about getting this AI out there for everyone to try. Because red teaming the chatbots is a thing that people on Reddit really enjoy doing.

Vedant Dwivedi:

Reddit is brutal.

Demetrios:

It's brutal, man. So what I think about though, when we take this idea and we apply it to real time voice agents is does it not slow down the voice agent because it has to go through these checks? And if not, how have you figured out a way to keep it so it sounds natural but while also being secure?

Vedant Dwivedi:

So the thing is that we use a framework known as Sigma. So Sigma is the one via which we make the voice AI agents and when we do it, we try to make as much possible on the distributed computing. So let's say I get my requests input, my user gave me the prompt. Now, at a runtime the guardrail is parallelly processed, then the input will go once it is being ticked, then it will go to my intel. It can be any API or it can be my GPU or it can be my model. The output of it will again go to my guardrail and then the deliverable whatever is the result will be getting to the user.

Vedant Dwivedi:

It does add a time and it does add a latency. But it is an absolute necessary mechanism if you want to have a safe and secure architecture. So how we try to reduce that gap is by applying it in a distributed cluster distributed mechanism so that it reduces the time, but a latency is merely of one or two seconds. So it does not affect a real time accuracy quite a lot.

Demetrios:

Have you seen that certain types of use cases are more inclined to be wanting a solution like this? That my mind automatically goes to finance or anything that's regulated.

Vedant Dwivedi:

So we are working with the manufacturing industries quite a lot. And what they want is that they want a real time voice agent which can tell them the insights insights of databases. And these databases are directly linked to their SAP. So it's a huge amount of data. So what they want is that I will give a voice instruction. Tell me what is the sales for the month July 2024 for a product known as the iron bars? How many iron bar were produced in manufacturing industry in July 2024? Now this has to go first, it has to be checked on the guardrails whether none of this query is absolutely a bullshit query. This has to be a real query to be asked to an intelligence MA intelligence machine makes the SQL queries and other queries based on top of the input logic.

Vedant Dwivedi:

And then this query goes to the db, it analyzes the data and that analyzed data is again converted into a string. And then this string goes to text to speech model and then text to speech model delivers the speech output to the real time customer. And all this we do in four seconds. So we try to reduce the gap and we try to make it as much secure as possible all in four seconds. And that is the real time use cases which we are implementing. Even for E commerce, we are doing the same. There is a new whole segment of order management system that we are trying right now and we are trying to do this exactly this over there as well.

Demetrios:

You spoke to me about three different types of ways you're attacking the problem. One being workflows, one being agents. Can you break down how you look at that?

Vedant Dwivedi:

So when we entered into the industry at that time we understood that the very first problem is data catalogs. And in order to make a strong data catalog, as in like every organization have their own scattered data. So we first recommend them a framework known as Builder via which they can do their data cataloging. And using this they get a real time copilot which manages all the data in the organization. Once the builder performs the operation of copilot, then we introduce another framework known as Sigma. Using Sigma they can then make their AI agents. But AI agents are very narrow, very focused and they tend to perform one specific problem at a time. Now when you want that there can be ton of AI agents and they should talk with one another, that is when you get introduced to flow.

Vedant Dwivedi:

Flow is a workflow where you can place one agent on the canvas, one agent talking to another agent, then this agent talking to another agent and then they all synchronously processing the data, talking among themselves and then providing you a real time insights. So this is the third product which we offer to them that is a flow and that's how we have segregated the problem and that's how we solve the problem.

Demetrios:

Now the data catalog piece is fascinating to me. I think you as a data scientist, you almost like have it in your DNA and in your blood. But for me, what exactly do you mean by that? And can you like ground that in a real world situation?

Vedant Dwivedi:

So let's say I come to you, I say I'll solve your data problems and I'll make a fantastic AI for you so that you can talk to your company's data all the time and you can get all the insights in real time. And you would say, I'm very happy, just do it for me. I mean it'll solve a lot of problem for me. But when I come and I start analyzing your data, I'll see that most of your data is on the SharePoint. Then some of your data is in Google Drive. Then you have your databases, you can have a SQL database, you can have NoSQL databases and also you are having some data which is in Excel, like so it can be a ton of pieces of data which is scattered around. So what we do is that we provide you Builder.

Vedant Dwivedi:

Builder is a tool. You just integrate it with your SQL databases, you integrate it with your NoSQL databases, you integrate it with your CSV file, flat system file. The builder learns understand and it catalog, it prepares its own metadata and it prepare its own catalog so that it can understand where the data is, how the data is, what is the quality of data how normalized the data is, what is the frequency of data, how often things are getting changed, how things are happening, all this understanding Builder will get. Once I get all the understanding of the data, I do understand how data is spread across organization, then I can make AI agents and once I have a lot of agents, I can make a workflow so that they can talk among one another.

Demetrios:

I guess that context that you can pass on to the agents is super helpful and relevant for them to be able to actually execute the task with success.

Vedant Dwivedi:

Because the real time problem is that it takes quite a long to implement one use case and then props and solve that particular use case. If you don't have a clean data in your enterprise. And that's a real time problem, that is the reason you would see that even after around two years of having the very greatest model, we don't have a real time application across all the industry because data is not cataloged. Data is still an issue in all the enterprises it's scattered across.

Demetrios:

So is Builder an abstraction of something like Apache Iceberg?

Vedant Dwivedi:

It's Apache Spark and Apache Dask running on an AI agent. That is what Builder is on the abstract. So it's parallel processing and doing the batch and performing the map and reduce to prepare the data catalogs. That is what Builder is. That's the abstract.

Demetrios:

That's super cool to think about. Now, what are some ways that you've actually seen these workflows be implemented? You mentioned before working with the industrial clients and they're looking to actually be able to talk with their data. Have you seen other ways that this has been useful? Especially because you're starting very early in the journey. You're not just coming in and saying, all right, we're going to slap on an agent at the end of this. That is customer facing. You're saying before we can get to this agent, we need to catalog the data, we gotta get it in tip top shape. Then we create these mini agents and then we can create the workflows from the mini agents.

Vedant Dwivedi:

So we actually proposed the solution after we did a lot of survey. We surveyed around 1000 enterprises client and in those surveys we understood that enterprises do want to have AI agents, but they can't just plug and play LLM to make an AI agent. So one way of actually saying that I'm having an AI agent is that you call a LLM API and then you do some function calling. But that does not solve any enterprise problems in real time because they have already had an automation like this before. So it doesn't add any real time value to them. For them a real AI agent is something which is self learning and something which is extremely reliable. Something which can be taken into production system.

Vedant Dwivedi:

And for that they always do a maker checker mechanism. You need to make your data and then check its data and then validate the data. Only then you go for an AI agent or else it can hallucinate, else it can give a wrong answer. So we have seen it with our real time clients. One of the best workflow which we made and we started from the builder, then we made Sigma and then we made a workflow for them is for the recruitment agency. They were having a scattered across data for all the resume. Then they were having all the data entry was done absolutely with a remote consultancy which they were having. And we just got it all together with help of a builder.

Vedant Dwivedi:

Then we prepared them four different kind of agents. One which perform all the resume analyzer for job descriptions. Second one which can do a conversational AI interviews. Like you just give the name of the candidate, the email ID and the phone number and the candidate will get a SMS and he will get a WhatsApp notification and also an email notification that you can just now talk and give your conversational AI interview. And this is where we use Deepgram quite a lot in this voice thing. And once that interview is done then we give another interview which is a video interview round where you just like we are having an interaction where we help the candidates to perform better but providing their facial insights as well. Like how nervous they were, how was their eye contact when they were answering the question, doing the whole facial recognition, understanding how.

Vedant Dwivedi:

How good their understanding is of the subject. We asked them the question and capture the facial expression. And the last one is the coding rounds where there is a a copilot and that co pilot assistant code alongs the real time candidate and then the copilot decides whether the candidate is performing coding task in a proper manner or not. And after doing all this four different agents have done the job for the same candidate. Your all the answers are augmented after the augmentation you get a common report which candidate performed better in all these four rounds. And so you can get the selection of the candidate based on this. So this is the whole workflow which we made for a consultancy. We started with the builder by cataloging the data.

Vedant Dwivedi:

Then we made four different kind of agents using Sigma and then we made a workflow which can run in a flow and can give you the best candidate at the end of the time. That is the HR workflow which we made for a client.

Demetrios:

Brilliant, dude. I think we'll end it there unless there's anything else that you want to say or talk about.

Vedant Dwivedi:

So the only thing is that we are really grateful for Deepgram for providing us with the benefits. I mean actually I'm grateful to you because you were the one who approved our startup brand. So thank you a lot for that. We've been using those credits really well and Deepgram is a still which we are using in all of our voice conversation. And thank you for the latest model as well. The latest model which was Nova 3 and then later it was I think the Enterprise was in the second version of the Enterprise one which you've just dropped. We are using that one also. So thank you so much for that.

Demetrios:

Incredible. Oh that's so cool. So good to hear.