Unlocking Business Insights: The Future of AI in Analytics with Chata.ai CEO Kelly Cherniwchan

0:28Welcome to the AI Chat Podcast. street from financial planning to innovation and product management. Kelly's diverse experience also extends to new company formation and risk management, making him a seasoned executive with a deep understanding of technology, finance, and business strategy. Welcome to the show today, Kelly. Thanks for having those. That was quite the background. Okay. I'll give you a secret, which I haven't told anyone yet, but how I get these bios, if someone doesn't send me their bio, which I never ask. So usually they don't is I actually will go to LinkedIn. I will just copy their entire experience page, give it a chat to PT and say, give me a good bio on it.

1:07I found it. It's got some really, it's great at putting it all together. And then I don't have to read through it on the first go around. So yeah, that's my secret in case anyone needs to make a quick bio. Good way to do it. There you go. So thank you so much for coming on the show today, Kelly. The first thing I want to kick this off and ask you, can you explain to everyone a little bit about your background and what got you interested and started in the tech space, AI space, right off of right, you know, at the beginning? Yeah, no, absolutely. So I actually was going to go into investment banking.

1:42So I was doing my master's. I was in business school, focused in financial derivatives and the mathematics around there. And that was at 2000. I graduated 2007. And that was kind of really when machine learning and those things were starting to pick up again. And so I ended up taking a job at a company creation group with a small seed fund. It was based out of the technology transfer office at UBC. And that's when one of the projects that came across the best was circle cardiovascular imaging for the very early stages of it. So I thought it had a lot of commercial potential. The market was still very niche, but sometimes you want to be the big fish in the small pond, knowing that that pond is going to get bigger.

2:22and jumped into things from there, started building the background more around, you know, understanding machine learning, how it works practically in the real world, both opportunities and challenges. And, you know, as a founding group, we saw where we needed to go. We knew that technology had to be built to get there. And we just all aligned on that whole piece. And we were actually one of the early adopters of convolutional neural nets. Like maybe if you look in the history books who had been one of the first medical imaging companies to actually use that for image segmentation and quantification as well.

2:57So yeah, that's kind of the background. The mathematics from the finance side gave, I'll say the mathematics background to be able to understand how neural networks work and understand a lot of those different pieces was an easy transition. And then you just have to learn software after that. Not that I code on a day-to-day basis at but at least I can hold a conversation with the engineers. Very cool. Very cool. So talk to us a little bit about, you know, you had this experience where you were working at that company. What made you decide to kind of jump on this opportunity with Shadda.ai? What made you kind of inspired to start this company?

3:38Yeah, no, I'll tell you the exact date. So we had just done a deal with a very large device imaging scanner device manufacturer. And so I'd been talking to one of the co-founders who's still the CEO of the company, Greg O 'Grodnik, just an amazing entrepreneur and really, really a good mentor. And so I was talking to him about an opportunity that I saw, something we struggled with at the last kind of few board meetings that we had, which was we would show these displays or these dashboards and reports in the board meeting. And what typically happened was there was about, I don't know, 10 questions that just got fired right off the bat.

4:20And a lot of the time we weren't able to answer those questions because they're pretty detailed questions. Right. And so then it's the typical, okay, well, we'll get back to you on that one because we got to go and run another report, go get something out of the database and return that information. And that's when I had started doing some work in computational linguistics, right? So this is the earlier days of NLP and thought that there is a potential that we could use natural language and actually have that translated to database query language in a robust way. And so that was, you know, we worked on a transition plan, wanted to start the company.

5:02Again, similar blueprint, knew where things were going to go, but that there was still technology that needed to be developed to get there. There's de-risking that needed to be done, everything else. And then so we began a multi-year research and development journey towards what we've built today. That's incredible. Talk to me a little bit about your process. So you want to start this company. I'm assuming you spoke with your mentor about this. You're going to have to leave your old company. Was this a spinoff company? This is a brand new thing. Did you go look for co-founders? Did you look for investors?

5:40What was the process of starting that? Yeah, it was not a subsidiary or anything. It was just my own thing. my own money to start as well. So it was just something that, you know, sometimes getting co-founders together, everybody has to be totally aligned. And I had like a very, very specific, a strong vision of kind of where I was expecting things to go. I talked to a few other people as well, but I don't think we're totally aligned right off the bat. So I got things going, outsourced a little bit of work. And then at that standpoint is when you started putting the team around so effectively not a not a founder off the bat but your first three to four people in the company are co-founders as far as i'm concerned so our cto our cto reggillette who is down in uh in las vegas as well at the ai4 conference uh you know he he's effectively a co-founder as well so uh yeah did that and again just aligned everything together made sure that we weren't just crazy and this was something that was actually achievable, not just, you know, as you see in the AI space nowadays, everybody wants this like magical AI that will do everything for them, which, you know, is still a pipe dream now.

7:01But where could we actually apply the technologies itself and really provide value in the real world scenarios, right? And so he came from the telecom background, handling billions and billions of rows of data coming into a system, how that gets managed. And, you know we align to who what type of market we wanted to go for and where we could make the the biggest impact very cool very cool talk to me a little bit about like explain i guess to the audience a little bit about exactly what chata does um and uh you know the problem that it's solving for customers yeah so the easiest way to think about what we do is it's we're just like google translate except instead of translating human to human language we translate human to database query language.

7:46Okay. And we do that to real two core technologies or two core pieces of IP. The first is the buzz term of today, which is generative AI. But we used to call it natural language generation. That's how you kind of test if the NLP space. Yeah, because it, I mean, it's just a different way. I mean, okay, architecturally, it's a little bit different. But the point is, is we use generative AI to build high quality training data because every single data warehouse database is unique in structure, it's unique in business logic, everything else. So it's not something you could just plug a pre-trained model into and it just works out of the box, okay?

8:29So we use generative AI to connect to, so we connect to a warehouse and we have our own technology which generates high quality training data, right? And the old adage, in AI is only as good as the training data that it's actually learned on. So we do that. And then we also have another system, which it's an automated training controller. So there's a lot of different things like automated hyperparameter tuning, as well as the ability to create new training data for itself as well. So we've built everything in a really strong pipeline to be able to rapidly build custom language models for our customers' proprietary databases and data warehouses.

9:12and then also deploy those models in their environment. So not one piece of proprietary data ever leaves the entire system. Nice. Nice. That's very cool. Wow, that's an incredible technology there. Talk to me a little bit about how you're helping in reducing data request backlogs for businesses. I understand that's probably a big feature, right? Yeah, that's the primary. like when we think about an ideal customer profile, there's really two key areas. One, it's an enterprise or even an ISV who's servicing their customers, but they have a big data request backlog, right? I mean, you can only put so much information in dashboards, but those dashboards are aggregated views that then get fed into there.

10:01So when you aggregate views, you've lost all that detailed information and the ability to join all that detailed information between tables and everything else. And so you constantly have these analysts having to go write custom, if the database query language is SQL for them, they're writing custom SQL reports. Business user has more requests. They see something in their dashboard and then there's something else they want to know, but they can't drill down or drill through those aggregated views. So it's this backlog like you talked about. So that's ICP part number one. The second one is personas that require on-demand access.

10:42So this is why the financial markets is a really important vertical we're working in. Retail is another one. You can't wait two weeks for information and get back. And in certain cases, it's two months with a lot of enterprises for custom reports. You need to make decisions quicker. right now on that backlog side that's the data analyst backlog what we've found and what the research has shown is there's the shadow backlog we call it the shadow backlog which is an order of magnitude more questions than non-technical business users have they never end up asking because it's going to you know the data analysts are way too busy it's going to take too long for them to get the data they just stop asking right i mean it's human nature you and i probably do it you know, do it all the time.

11:26And so that, that shadow backlog, when we talk about tightening economic conditions and everything else, that is where you drive the ROI from the data warehouse and digital transformation investments that these enterprises have actually made. And so that from a positioning standpoint is, is where we go. And then between you and me, the analysts do, well, I guess there's probably some analysts listening as well, but they will resonate with the fact that there's nothing more boring than writing basic, you know, SQL statements to grab very simple reports for business people. You want to be working on advanced analysis.

12:05And in certain organizations, the data analysts are also in the machine learning, well, doing machine learning project. That's the stuff that adds ultimate value. So we're kind of there to clean up all that, Kate. You have data requests. We'll handle the data requested volume. They're really complex analytics stuff. That's what the data teams do. They love it. We're aligned from there and move forward with the success. Very cool. Yeah, that makes a lot of sense. And you mentioned those analysts there. So I believe your solution, it's really designed for non-technical business users, but also it sounds like analysts really love it.

12:42So what features have kind of made it appeal to both of those demographics? Well, the main, I mean, yeah, the main features, there's a lot of data analysts out there who are incredibly proficient with database query language buildings. So in that case, we can help from there. But the system is an API first. Like the system is more of a developer platform than anything else. Okay, I like it. It can be embedded anywhere, everything else. So there's things like Microsoft Excel plugins and all those. So data analysts who have models built within Microsoft Excel, and there's a lot of them, right? Their pain point is they just can't just get the source data into their Excel file.

13:24So the plugin allows them to query the warehouse, which then brings the data directly into Excel. And all they have to do is highlight the cells they want for their predictive model from there. Primarily, though, it is the non-technical user. They're basically just stuck. But the analysts like it because, again, it takes that basic reporting backlog off of their desks, right? And when you think about an organization who says, look, we have to do this faster. We have to get information to our non-technical decision makers faster. The only way to solve that problem without kind of solutions that we work on is to hire more people, right?

14:07And that's the only way to do it. So we're not there to say, hey, now you don't need all these people. We're there to say to eliminate this backlog, you don't have to bring on a bunch more headcount for something that technology is actually. Yes, that makes a lot of sense. And I can see some serious benefit to businesses from that. So I've noticed you have mentioned that you've done some partnerships. You've done partnerships, some big names, Microsoft, Snowblake, a handful of others. Talk to me a little bit about how you initiated those conversations with those companies, what those partnerships look like today, and maybe how those are helping you and your customers.

14:47Yeah, so partnering with Snowflake is more of a new thing, but they're a really good, have strong partner potential because they are a very well-respected brand name. and and to be honest with you the majority of our projects so far are actually snowflake data warehouses right uh there's some you're seeing a big big increase in databricks as well yeah but a lot of them are really around a snowflake where they can take their sap data and their and their salesforce data and put it together in snowflake and then now you have the warehouse that that uh that has all that information that you can uh query from from there uh microsoft's is very much a commercial partnership, which is you have these big consumption amounts that organizations pay for.

15:44They want to look for value-add systems and tools that they can establish within the enterprise to drive more ROI on their investments from that standpoint. So we are part of Microsoft Transact. So a purchase of our system can work through those consumption commitments that are becoming more and more important on a day-to-day basis. And we do use, you know, we use lots of systems where multi-cloud we can deploy in a bunch of different clouds, but we definitely see, especially Azure, lots in the enterprise. And so we've been able to leverage some of those uh tools and systems with model deployments and everything else so we have some strong technical ties there but microsoft has been a really really good partner uh you know opened some some good doors for us uh which is always as you know in in startup land that's one of the most important things totally 100 percent um so talk to me a little bit about about like who your ideal customer is for this product is this something that's primarily used like you've done partnerships in these big companies is something primarily used by large corporations do small businesses use this who's your you know who's your your typical customer for this yeah now enterprises for sure um we had looked when we were first de-risking the model or de-risking the actual technology we we pre-trained some models for the quickbooks data warehouse and the zero data warehouse and actually Stripe as well.

17:22All right. Cause they published their object model, everything else. And, and as part of that, to get the most possible users and the most possible queries through the system, you deploy that, right? So we created an app, put an app around it and deployed it from there. The system did extremely well, but the distribution channels and getting that out to the small businesses, it, it normally goes to resellers at the end of the day. And so the, I wouldn't say the product market fit, but I would say the market distribution fit was not super strong. But that was never where we were intending to go.

17:56Verizon tended to go to the enterprise, right, as a key validation step, and then also be able to deploy the entire system out so that do-it-yourself people could actually use our technology to build their own proprietary language models internally within the system. So we started that way, but we absolutely target enterprises with those two factors I was mentioning before, with data analyst backlog and the need for on-demand that data access. And right now we're in the nice spot that we're getting so much inbound demand on this, which is, you know, well, you know, I think when we were talking at the AI4 booth, you saw a bunch of other kind of people even came by as well.

18:43And that's because of the advent of ChatGPT and now even some of the way people are finding us on searches is via that as well, because we don't do all of the things that these autoregressive LLMs, these pre-trained ones do. We have a very specific area, which is database query generation, which those systems actually don't do very well and are highly risky to do that. right so we formed in a nice little area in there and so in the enterprise as you know chat gpt and these other uh these other big llms they're being adopted they're adding tons of value all those other pieces so we're driving as part of that wave uh and we are getting lots of inbound because of that that's very cool buzz that's being created that that's amazing and i think it's really interesting too like you obviously are not brand new in the space you know you've been looking in space for a long time working at the space at a long time for you what do you think like one of the biggest differentiators between you know chatta is versus maybe some of these new players that are trying to come into the space today where do you see uh you know the most value and differentiation yeah so so a lot of the new like a lot of the new entrants trying to use the llms they're actually not doing anything different than the business intelligence tools have done for the last year and a half, two years.

20:05So there's like Tableau-esque data, Sisense has a system, a few of the other ones. But when you look into the docs, they have the language model or they have the interface based on that view, right? Which typically is limited to the number of columns, maybe 50 or 20 columns. Power BI is the same thing. So those new systems coming in, what they do with the LLMs is they create a view and then prompt the LLM with a view, okay? So you don't have to worry about joins. complex subqueries aren't there there's a whole bunch of things that just aren't supported which like you said we've been doing this for a long time there's a lot of real difficult practical things when you're building database queries but the biggest challenge is when you're using again we use generative ai on the training data side of things if you use it on the inference side of things you need to be absolutely consistent every time because you have a bad join in a database query and you can bring that entire system down right so we all know because we've all worked with these things and work with the big l11s on a day-to-day basis they don't need to be 100 perfect nor are they we just need to get 90 95 of the way there we can take that stuff copy paste you know change it however we see fit right and then we go from there but generating database queries is very different you need to have the system designed and trained for that particular structure so even a simple question like um you know uh change in monthly sales for product xyz and region xyz during campaign xyz like that detailed information that we work with that exact question in the exact same company type would be two fundamentally different database queries.

21:59Because that's dependent on the thing. And you can't prompt an LLM with an entire warehouse schema, right? Right. It's lots. There's conflicts. You just don't get it. So a lot of the people that have been trying this area now, and to be totally frank with you, we're not the first choice in a lot of these cases. A lot of these large organizations, they try, they've tried to go with this. And then they come to us and I can't name any names, but actually a few in the last two weeks have come to be like, we've tried this. It just doesn't work. It's not consistent. We've tried all the guardrails and everything else we can.

22:38It's not going to do the database stuff. It does our document search and everything else we need, which is really, really what it's quite strong at. But we need a different solution. Can you integrate in this type of entire environment together? And that's when we say absolutely. The API and it just, you call our API when you want a query that needs information from a data warehouse. That query will make the joins, do the calculations, anything else that's actually required. Very cool. Very cool. So I can see some massive value that this is obviously adding. but I'm super curious you've been at this for a while and but because this there's this whole new kind of wave of hype and interest and a lot of companies as you say are kind of reaching out to you now what what's kind of new or what's next on the horizon for chata.ai are there any upcoming features or improvements that you're you guys are looking at that you can share with us yeah I mean we're always trying to improve so I'll give you I'll give you a new technology that we're just rolling out now okay so uh we can in the past we were able to train a custom language model in about depending upon the size of the warehouse in about 48 hours last kind of what what the training would actually take okay so that's what it takes from there now and then you iterate on the model because people want to add new columns or new tables or anything like this that's the way so the thing's set up for for iteration for sure so that 48 hours now with this new rollout we've actually dropped that to one hour wow so we can have a language model trained in one hour the thing nobody talks about in the space is how expensive it is to have large language models actually run let alone train them yeah right so now all of a sudden a team can iterate on models and they got to wait one hour for it to actually trade to be able to use right away.

24:38Okay? So I can't go too much deep into the technology, but it's about, you know, instead of using a sledgehammer to hit every single nail in, right, you can have specialty hammers that can work together to still get the job done, right? As long as they can work together, right? So, you know, you pointed out before, we've been working on this a long time. We've gone through lots of different iterations, tried lots of different things. The last system we had has been working incredibly well. This is a way that we make things much quicker, much more economic than everything else. And then the second piece would be around, there's a lot of information that can be generated from these questions that people are actually asking.

25:21right and so we have a really really strong research direction and understanding that trying to help people reason through certain analysis uh and the and those types of pieces as well so we have a really robust uh r &d plan uh our kind of core system now is is yeah i mean now now is the time so now we've hit the market you know three years ago where we're We were early to market. People quite weren't ready for something like this. But again, because of our friends at OpenAI and Microsoft and now everybody else, they're pushing the conversational. Because the chatbot's kind of wrecked it for everybody.

26:00Right. Everyone is like, conversationally, I don't want to touch on the tentacle. And so now the auto-men, oh, this stuff actually works. And so now things are pushing forward. But there is a very bright future in that area without it. Sure. And that's the reason why, you know, we're we're raising it 10 million U.S. round because we have this pipeline to execute on, which is absolutely amazing. Very, very good brands. But we have to keep that R &D engine pumping because we got a lot of good, really strong ideas that we were ahead in a lot of areas. Right. So, you know, when GPT-4 came out, everybody's like, oh, that's not one model.

26:39That's actually a bunch of different models that are kind of working together. that's exactly the architecture yeah right now no don't get me wrong i'm not saying that like ours is that massive and can do all the things there but for our little our little corner of the earth in the natural language to database query language that's the philosophy that that we took and our team's very brilliant they're very strong and they come up with a lot of really interesting ideas and as a startup we can execute quite quite quick on those ideas for sure that's very cool and one thing i do want to highlight that you said there because i think it's really interesting you know i know you can't go in too much into the tech but you you know you mentioned not using a sledgehammer to do something but kind of smaller um more accurate hammers i think you guys are absolutely on the right track there be um a bunch of research that we've seen out of google deep mind for example um they they just trained a new chess model where essentially it beat alpha zero because what it did is they spun up 16 individual like chess model players and they all trained them with slightly different styles and then instead of having one kind of major sledgehammer as it were to play the game of chess they have the 16 all going against each or like collaborating and deciding which of the 16 had the best move the same thing when opening that's exactly it yes the ability that these different models can pass information back and forth you You're exactly correct.

28:05That is really cool. And in addition, when OpenAI's GPT-4 model weights were leaked, it was found that they have 16 experts within GPT-4 that when you have your queries, it collaborates with all that. So I think 100 % you guys are on the right direction. I'm super excited about your$10 million round. I think you guys are... The companies that I see successful right now, raising successfully and really killing it in this industry, are companies like yourself that have been in this for a long time. It's maybe not the newest thing, but it's something that has a really solid tech background. And now you're taking this to the next level.

28:41You have the perfect position to be a really powerful launchpad, pushing forward your technology. And the tech's already there. Now it's time to take it to the next level. And so, yeah, I'm really excited for you guys. I think you'll be incredibly successful in that. Listen, Kelly, thank you so much for coming on the podcast today, for sharing your insights. Really excited about what you guys are building over at Chata AI. If people want to try out Chata or if they want to get in contact with you, what is the best way for them to do that? Yeah, I mean, they can go to our website or they can just reach out directly to me as well.

29:15I can leave my email contact information with you when you post that. And yeah, I'm happy to talk to potential partners, investors, whomever. or just people are generally interested in the area is always great conversations for sure. Wonderful. I'll leave that in the description for the show notes. So to the listeners, thank you so much for tuning into the AI Chat Podcast. Make sure to rate us wherever you listen to your podcasts and have a wonderful rest of your day.