Transcript

Announcer: Hello, and welcome to Screaming in the Cloud with your host, Chief Cloud Economist at The Duckbill Group, Corey Quinn. This weekly show features conversations with people doing interesting work in the world of cloud, thoughtful commentary on the state of the technical world, and ridiculous titles for which Corey refuses to apologize. This is Screaming in the Cloud.

Corey: This episode is sponsored in part by our friends at Chronosphere. Tired of observability costs going up every year without getting additional value? Or being locked into a vendor due to proprietary data collection, querying, and visualization? Modern-day, containerized environments require a new kind of observability technology that accounts for the massive increase in scale and attendant cost of data. With Chronosphere, choose where and how your data is routed and stored, query it easily, and get better context and control. 100% open-source compatibility means that no matter what your setup is, they can help. Learn how Chronosphere provides complete and real-time insight into ECS, EKS, and your microservices, wherever they may be at snark.cloud/chronosphere that’s snark.cloud/chronosphere.

Corey: This episode is brought to you in part by our friends at Veeam. Do you care about backups? Of course you don’t. Nobody cares about backups. Stop lying to yourselves! You care about restores, usually right after you didn’t care enough about backups. If you’re tired of the vulnerabilities, costs, and slow recoveries when using snapshots to restore your data, assuming you even have them at all living in AWS-land, there is an alternative for you. Check out Veeam, that's V-E-E-A-M for secure, zero-fuss AWS backup that won’t leave you high and dry when it’s time to restore. Stop taking chances with your data. Talk to Veeam. My thanks to them for sponsoring this ridiculous podcast.

Corey: Welcome to Screaming in the Cloud. I’m Corey Quinn. Today’s promoted guest episode is brought to us by our friends at Pinecone and they have given their VP of Engineering and R&D over to suffer my various sling and arrows, Ram Sriharsha. Ram, thank you for joining me.

Ram: Corey, great to be here. Thanks for having me.

Corey: So, I was immediately intrigued when I wound up seeing your website, pinecone.io because it says right at the top—at least as of this recording—in bold text, “The Vector Database.” And if there’s one thing that I love, it is using things that are not designed to be databases as databases, or inappropriately referring to things—be they JSON files or senior engineers—as databases as well. What is a vector database?

Ram: That’s a great question. And we do use this term correctly, I think. You can think of customers of Pinecone as having all the data management problems that they have with traditional databases; the main difference is twofold. One is there is a new data type, which is vectors. Vectors, you can think of them as arrays of floats, floating point numbers, and there is a new pattern of use cases, which is search.

And what you’re trying to do in vector search is you’re looking for the nearest, the closest vectors to a given query. So, these two things fundamentally put a lot of stress on traditional databases. So, it’s not like you can take a traditional database and make it into a vector database. That is why we coined this term vector database and we are building a new type of vector database. But fundamentally, it has all the database challenges on a new type of data and a new query pattern.

Corey: Can you give me an example of what, I guess, an idealized use case would be of what the data set might look like and what sort of problem you would have in a vector database would solve?

Ram: A very great question. So, one interesting thing is there’s many, many use cases. I’ll just pick the most natural one which is text search. So, if you’re familiar with the Elastic or any other traditional text search engines, you have pieces of text, you index them, and the indexing that you do is traditionally an inverted index, and then you search over this text. And what this sort of search engine does is it matches for keywords.

So, if it finds a keyword match between your query and your corpus, it’s going to retrieve the relevant documents. And this is what we call text search, right, or keyword search. You can do something similar with technologies like Pinecone, but what you do here is instead of searching our text, you’re searching our vectors. Now, where do these vectors come from? They come from taking deep-learning models, running your text through them, and these generate these things called vector embeddings.

And now, you’re taking a query as well, running them to deep-learning models, generating these query embeddings, and looking for the closest record embeddings in your corpus that are similar to the query embeddings. This notion of proximity in this space of vectors tells you something about semantic similarity between the query and the text. So suddenly, you’re going beyond keyword search into semantic similarity. An example is if you had a whole lot of text data, and maybe you were looking for ‘soda,’ and you were doing keyword search. Keyword search will only match on variations of soda. It will never match ‘Coca-Cola’ because Coca-Cola and soda have nothing to do with each other.

Corey: Or Pepsi, or pop, as they say in the American Midwest.

Ram: Exactly.

Corey: Yeah.

Ram: Exactly. However, semantic search engines can actually match the two because they’re matching for intent, right? If they find in this piece of text, enough intent to suggest that soda and Coca-Cola or Pepsi or pop are related to each other, they will actually match those and score them higher. And you’re very likely to retrieve those sort of candidates that traditional search engines simply cannot. So, this is a canonical example, what’s called semantic search, and it’s known to be done better by these other vector search engines. There are also other examples in say, image search. Just if you’re looking for near duplicate images, you can’t even do this today without a technology like vector search.

Corey: What is the, I guess, translation or conversion process of existing dataset into something that a vector database could use? Because you mentioned it was an array of floats was the natural vector datatype. I don’t think I’ve ever seen even the most arcane markdown implementation that expected people to wind up writing in arrays of floats. What does that look like? How do you wind up, I guess, internalizing or ingesting existing bodies of text for your example use case?

Ram: Yeah, this is a very great question. This used to be a very hard problem and what has happened over the last several years in deep-learning literature, as well as in deep-learning as a field itself, is that there have been these large, publicly trained models, examples will be OpenAI, examples will be the models that are available in Hugging Face like Cohere, and a large number of these companies have come forward with very well trained models through which you can pass pieces of text and get these vectors. So, you no longer have to actually train these sort of models, you don’t have to really have the expertise to deeply figured out how to take pieces of text and build these embedding models. What you can do is just take a stock model, if you’re familiar with OpenAI, you can just go to OpenAIs homepage and pick a model that works for you, Hugging Face models, and so on. There’s a lot of literature to help you do this.

Sophisticated customers can also do something called fine-tuning, which is built on top of these models to fine-tune for their use cases. The technology is out there already, there’s a lot of documentation available. Even Pinecone’s website has plenty of documentation to do this. Customers of Pinecone do this [unintelligible 00:07:45], which is they take piece of text, run them through either these pre-trained models or through fine-tuned models, get the series of floats which represent them, vector embeddings, and then send it to us. So, that’s the workflow. The workflow is basically a machine-learning pipeline that either takes a pre-trained model, passes them through these pieces of text or images or what have you, or actually has a fine-tuning step in it.

Corey: Is that ingest process something that not only benefits from but also requires the use of a GPU or something similar to that to wind up doing the in-depth, very specific type of expensive math for data ingestion?

Ram: Yes, very often these run on GPUs. Sometimes, depending on budget, you may have compressed models or smaller models that run on CPUs, but most often they do run on GPUs, most often, we actually find people make just API calls to services that do this for them. So, very often, people are actually not deploying these GPU models themselves, they are maybe making a call to Hugging Face’s service, or to OpenAI’s service, and so on. And by the way, these companies also democratized this quite a bit. It was much, much harder to do this before they came around.

Corey: Oh, yeah. I mean, I’m reminded of the old XKCD comic from years ago, which was, “Okay, I want to give you a picture. And I want you to tell me it was taken within the boundaries of a national park.” Like, “Sure. Easy enough. Geolocation information is attached. It’ll take me two hours.” “Cool. And I also want you to tell me if it’s a picture of a bird.” “Okay, that’ll take five years and a research team.”

And sure enough, now we can basically do that. The future is now and it’s kind of wild to see that unfolding in a human perceivable timespan on these things. But I guess my question now is, so that is what a vector database does? What does Pinecone specifically do? It turns out that as much as I wish it were otherwise, not a lot of companies are founded on, “Well, we have this really neat technology, so we’re just going to be here, well, in a foundational sense to wind up ensuring the uptake of that technology.” No, no, there’s usually a monetization model in there somewhere. Where does Pinecone start, where does it stop, and how does it differentiate itself from typical vector databases? If such a thing could be said to exist yet.

Ram: Such a thing doesn’t exist yet. We were the first vector database, so in a sense, building this infrastructure, scaling it, and making it easy for people to operate it in a SaaS fashion is our primary core product offering. On top of that, this very recently started also enabling people who have who actually have raw text to not just be able to get value from these vector search engines and so on, but also be able to take advantage of traditional what we call keyword search or sparse retrieval and do a combined search better, in Pinecone. So, there’s value-add on top of this that we do, but I would say the core of it is building a SaaS managed platform that allows people to actually easily store as data, scale it, query it in a way that’s very hands off and doesn’t require a lot of tuning or operational burden on their side. This is, like, our core value proposition.

Corey: Got it. There’s something to be said for making something accessible when previously it had only really been available to people who completed the Hello World tutorial—which generally resembled a doctorate at Berkeley or Waterloo or somewhere else—and turn it into something that’s fundamentally, click the button. Where on that, I guess, a spectrum of evolution do you find that Pinecone is today?

Ram: Yeah. So, you know, prior to Pinecone, we didn’t really have this notion of a vector database. For several years, we’ve had libraries that are really good that you can pre-train on your embeddings, generate this thing called an index, and then you can search over that index. There is still a lot of work to be done even to deploy that and scale it and operate it in production and so on. Even that was not being, kind of, offered as a managed service before.

What Pinecone does which is novel, is you no longer have to have this pre-training be done by somebody, you no longer have to worry about when to retrain your indexes, what to do when you have new data, what to do when there is deletions, updates, and the usual data management operations. You can just think of this is, like, a database that you just throw your data in. It does all the right things for you, you just worry about querying. This has never existed before, right? This is—it’s not even like we are trying to make the operational part of something easier. It is that we are offering something that hasn’t existed before, at the same time, making it operationally simple.

So, we’re solving two problems, which is we building a better database that hasn’t existed before. So, if you really had this sort of data management problems and you wanted to build an index that was fresh that you didn’t have to super manually tune for your own use cases, that simply couldn’t have been done before. But at the same time, we are doing all of this in a cloud-native fashion; it’s easy for you to just operate and not worry about.

Corey: You’ve said that this hasn’t really been done before, but this does sound like it is more than passingly familiar specifically to the idea of nearest neighbor search, which has been around since the ’70s in a bunch of different ways. So, how is it different? And let me of course, ask my follow-up to that right now: why is this even an interesting problem to start exploring?

Ram: This is a great question. First of all, nearest neighbor search is one of the oldest forms of machine learning. It’s been known for decades. There’s a lot of literature out there, there are a lot of great libraries as I mentioned in the passing before. All of these problems have primarily focused on static corpuses. So basically, you have a set of some amount of data, you want to create an index out of it, and you want to query it.

A lot of literature has focused on this problem. Even there, once you go from small number of dimensions to large number of dimensions, things become computationally far more challenging. So, traditional nearest neighbor search actually doesn’t scale very well. What do I mean by large number of dimensions? Today, deep-learning models that produce image representations typically operate in 2048 dimensions of photos [unintelligible 00:13:38] dimensions. Some of the OpenAI models are even 10,000 dimensional and above. So, these are very, very large dimensions.

Most of the literature prior to maybe even less than ten years back has focused on less than ten dimensions. So, it’s like a scale apart in dealing with small dimensional data versus large dimensional data. But even as of a couple of years back, there hasn’t been enough, if any, focus on what happens when your data rapidly evolves. For example, what happens when people add new data? What happens if people delete some data? What happens if your vectors get updated? These aren’t just theoretical problems; they happen all the time. Customers of ours face this all the time.

In fact, the classic example is in recommendation systems where user preferences change all the time, right, and you want to adapt to that, which means your user vectors change constantly. When even these sort of things change constantly, you want your index to reflect it because you want your queries to catch on to the most recent data. [unintelligible 00:14:33] have to reflect the recency of your data. This is a solved problem for traditional databases. Relational databases are great at solving this problem. A lot of work has been done for decades to solve this problem really well.

This is a fundamentally hard problem for vector databases and that’s one of the core focus areas [unintelligible 00:14:48] painful. Another problem that is hard for these sort of databases is simple things like filtering. For example, you have a corpus of say product images and you want to only look at images that maybe are for the Fall shopping line, right? Seems like a very natural query. Again, databases have known and solved this problem for many, many years.

The moment you do nearest neighbor search with these sort of constraints, it’s a hard problem. So, it’s just the fact that nearest neighbor search and lots of research in this area has simply not focused on what happens to that, so those are of techniques when combined with data management challenges, filtering, and all the traditional challenges of a database. So, when you start doing that you enter a very novel area to begin with.

Corey: This episode is sponsored in part by our friends at Redis, the company behind the incredibly popular open-source database. If you’re tired of managing open-source Redis on your own, or if you are looking to go beyond just caching and unlocking your data’s full potential, these folks have you covered. Redis Enterprise is the go-to managed Redis service that allows you to reimagine how your geo-distributed applications process, deliver and store data. To learn more from the experts in Redis how to be real-time, right now, from anywhere, visit snark.cloud/redis. That’s snark dot cloud slash R-E-D-I-S.

Corey: So, where’s this space going, I guess is sort of the dangerous but inevitable question I have to ask. Because whenever you talk to someone who is involved in a very early stage of what is potentially a transformative idea, it’s almost indistinguishable from someone who is whatever the polite term for being wrapped around their own axle is, in a technological sense. It’s almost a form of reverse Schneier’s Law of anyone can create an encryption algorithm that they themselves cannot break. So, the possibility that this may come back to bite us in the future if it turns out that this is not potentially the revelation that you see it as, where do you see the future of this going?

Ram: Really great question. The way I think about it is, and the reason why I keep going back to databases and these sort of ideas is, we have a really great way to deal with structured data and structured queries, right? This is the evolution of the last maybe 40, 50 years is to come up with relational databases, come up with SQL engines, come up with scalable ways of running structured queries on large amounts of data. What I feel like this sort of technology does is it takes it to the next level, which is you can actually ask unstructured questions on unstructured data, right? So, even the couple of examples we just talked about, doing near duplicate detection of images, that’s a very unstructured question. What does it even mean to say that two images are nearly duplicate of each other? I couldn’t even phrase it as kind of a concrete thing. I certainly cannot write a SQL statement for it, but I cannot even phrase it properly.

With these sort of technologies, with the vector embeddings, with deep learning and so on, you can actually mathematically phrase it, right? The mathematical phrasing is very simple once you have the right representation that understands your image as a vector. Two images are nearly duplicate if they are close enough in the space of vectors. Suddenly you’ve taken a problem that was even hard to express, let alone compute, made it precise to express, precise to compute. This is going to happen not just for images, not just for semantic search, it’s going to happen for all sorts of unstructured data, whether it’s time series, where it’s anomaly detection, whether it’s security analytics, and so on.

I actually think that fundamentally, a lot of fields are going to get disrupted by this sort of way of thinking about things. We are just scratching the surface here with semantic search, in my opinion.

Corey: What is I guess your barometer for success? I mean, if I could take a very cynical point of view on this, it’s, “Oh, well, whenever there’s a managed vector database offering from AWS.” They’ll probably call it Amazon Basics Vector or something like that. Well, that is a—it used to be a snarky observation that, “Oh, we’re not competing, we’re just validating their market.” Lately, with some of their competitive database offerings, there’s a lot more truth to that than I suspect AWS would like.

Their offerings are nowhere near as robust as what they pretend to be competing against. How far away do you think we are from the larger cloud providers starting to say, “Ah, we got the sense there was money in here, so we’re launching an entire service around this?”

Ram: Yeah. I mean, this is a—first of all, this is a great question. There’s always something that’s constantly, things that any innovator or disrupter has to be thinking about, especially these days. I would say that having a multi-year head, start in the use cases, in thinking about how this system should even look, what sort of use cases should it [unintelligible 00:19:34], what the operating points for the [unintelligible 00:19:37] database even look like, and how to build something that’s cloud-native and scalable, is very hard to replicate. Meaning if you look at what we have already done and kind of tried to base the architecture of that, you’re probably already a couple of years behind us in terms of just where we are at, right, not just in the architecture, but also in the use cases in where this is evolving forward.

That said, I think it is, for all of these companies—and I would put—for example, Snowflake is a great example of this, which is Snowflake needn’t have existed if Redshift had done a phenomenal job of being cloud-native, right, and kind of done that before Snowflake did it. In hindsight, it seems like it’s obvious, but when Snowflake did this, it wasn’t obvious that that’s where everything was headed. And Snowflake built something that’s very technologically innovative, in a sense that it’s even now hard to replicate. Plus, it takes a long time to replicate something like that. I think that’s where we are at.

If Pinecone does its job really well and if we simply execute efficiently, it’s very hard to replicate that. So, I’m not super worried about cloud providers, to be honest, in this space, I’m more worried about our execution.

Corey: If it helps anything, I’m not very deep into your specific area of the world, obviously, but I am optimistic when I hear people say things like that. Whenever I find folks who are relatively early along in their technological journey being very concerned about oh, the large cloud provider is going to come crashing in, it feels on some level like their perspective is that they have one weird trick, and they were able to crack that, but they have no defensive mode because once someone else figures out the trick, well, okay, now we’re done. The idea of sustained and lasting innovation in a space, I think, is the more defensible position to take, with the counterargument, of course, that that’s a lot harder to find.

Ram: Absolutely. And I think for technologies like this, that’s the only solution, which is, if you really want to avoid being disrupted by cloud providers, I think that’s the way to go.

Corey: I want to talk a little bit about your own background. Before you wound up as the VP of R&D over at Pinecone, you were in a bunch of similar… I guess, similar styled roles—if we’ll call it that—at Yahoo, Databricks, and Splunk. I’m curious as to what your experience in those companies wound up impressing on you that made you say, “Ah, that’s great and all, but you know what’s next? That’s right, vector databases.” And off, you went to Pinecone. What did you see?

Ram: So, first of all, in was some way or the other, I have been involved in machine learning and systems and the intersection of these two for maybe the last decade-and-a-half. So, it’s always been something, like, in the in between the two and that’s been personally exciting to me. So, I’m kind of very excited by trying to think about new type of databases, new type of data platforms that really leverages machine learning and data. This has been personally exciting to me. I obviously learned very different things from different companies.

I would say that Yahoo was just the learning in cloud to begin with because prior to joining Yahoo, I wasn’t familiar with Silicon Valley cloud companies at that scale and Yahoo is a big company and there’s a lot to learn from there. It was also my first introduction to Hadoop, Spark, and even machine learning where I really got into machine learning at scale, in online advertising and areas like that, which was a massive scale. And I got into that in Yahoo, and it was personally exciting to me because there’s very few opportunities where you can work on machine learning at that scale, right?

Databricks was very exciting to me because it was an earlier-stage company than I had been at before. Extremely well run and I learned a lot from Databricks, just the team, the culture, the focus on innovation, and the focus on product thinking. I joined Databricks as a product manager. I hadn’t played the product manager hat before that, so it was very much a learning experience for me and I think I learned from some of the best in that area. And even at Pinecone, I carry that forward, which is think about how my learnings at Databricks informs how we should be thinking about products at Pinecone, and so on. So, I think I learned—if I had to pick one company I learned a lot from, I would say, it’s Databricks. The most [unintelligible 00:23:50].

Corey: I would also like to point out, normally when people say, “Oh, the one company I’ve learned the most from,” and they pick one of them out of their history, it’s invariably the most recent one, but you left there in 2018—

Ram: Yeah.

Corey: —then went to go spend the next three years over at Splunk, where you were a Senior Principal, Scientist, a Senior Director and Head of Machine-Learning, and then you decided, okay, that’s enough hard work. You’re going to do something easier and be the VP of Engineering, which is just wild at a company of that scale.

Ram: Yeah. At Splunk, I learned a lot about management. I think managing large teams, managing multiple different teams, while working on very different areas is something I learned at Splunk. You know, I was at this point in my career when I was right around trying to start my own company. Basically, I was at a point where I’d taken enough learnings and I really wanted to do something myself.

That’s when Edo and I—you know, the CEO of Pinecone—and I started talking. And we had worked together for many years, and we started working together at Yahoo. We kept in touch with each other. And we started talking about the sort of problems that I was excited about working on and then I came to realize what he was working on and what Pinecone was doing. And we thought it was a very good fit for the two of us to work together.

So, that is kind of how it happened. It sort of happened by chance, as many things do in Silicon Valley, where a lot of things just happen by network and chance. That’s what happened in my case. I was just thinking of starting my own company at the time when just a chance encounter with Edo led me to Pinecone.

Corey: It feels from my admittedly uninformed perspective, that a lot of what you’re doing right now in the vector database area, it feels on some level, like it follows the trajectory of machine learning, in that for a long time, the only people really excited about it were either sci-fi authors or folks who had trouble explaining it to someone without a degree in higher math. And then it turned into—a couple of big stories from the mid-2010s stick out at me when we’ve been people were trying to sell this to me in a variety of different ways. One of them was, “Oh, yeah, if you’re a giant credit card processing company and trying to detect fraud with this kind of transaction volume—” it’s, yeah, there are maybe three companies in the world that fall into that exact category. The other was WeWork where they did a lot of computer vision work. And they used this to determine that at certain times of day there was congestion in certain parts of the buildings and that this was best addressed by hiring a second barista. Which distilled down to, “Wait a minute, you’re telling me that you spent how much money on machine-learning and advanced analyses and data scientists and the rest have figured out that people like to drink coffee in the morning?” Like, that is a little on the ridiculous side.

Now, I think that it is past the time for skepticism around machine learning when you can go to a website and type in a description of something and it paints a picture of the thing you just described. Or you can show it a picture and it describes what is in that picture fairly accurately. At this point, the only people who are skeptics, from my position on this, seem to be holding out for some sort of either next-generation miracle or are just being bloody-minded. Do you think that there’s a tipping point for vector search where it’s going to become blindingly obvious to, if not the mass market, at least more run-of-the-mill, more prosaic level of engineer that haven’t specialized in this?

Ram: Yeah. It’s already, frankly, started happening. So, two years back, I wouldn’t have suspected this fast of an adoption for this new of technology from this varied number of use cases. I just wouldn’t have suspected it because I, you know, I still thought, it’s going to take some time for this field to mature and, kind of, everybody to really start taking advantage of this. This has happened much faster than even I assumed.

So, to some extent, it’s already happening. A lot of it is because the barrier to entry is quite low right now, right? So, it’s very easy and cost-effective for people to create these embeddings. There is a lot of documentation out there, things are getting easier and easier, day by day. Some of it is by Pinecone itself, by a lot of work we do. Some of it is by, like, companies that I mentioned before who are building better and better models, making it easier and easier for people to take these machine-learning models and use them without having to even fine-tune anything.

And as technologies like Pinecone really mature and dramatically become cost-effective, the barrier to entry is very low. So, what we tend to see people do, it’s not so much about confidence in this new technology; it is connecting something simple that I need this sort of value out of, and find the least critical path or the simplest way to get going on this sort of technology. And as long as it can make that barrier to entry very small and make this cost-effective and easy for people to explore, this is going to start exploding. And that’s what we are seeing. And a lot of Pinecone’s focus has been on ease-of-use, in simplicity in connecting the zero-to-one journey for precisely this reason. Because not only do we strongly believe in the value of this technology, it’s becoming more and more obvious to the broader community as well. The remaining work to be done is just the ease of use and making things cost-effective. And cost-effectiveness is also what the focus on a lot. Like, this technology can be even more cost-effective than it is today.

Corey: I think that it is one of those never-mistaken ideas to wind up making something more accessible to folks than keeping it in a relatively rarefied environment. We take a look throughout the history of computing in general and cloud in particular, were formerly very hard things have largely been reduced down to click the button. Yes, yes, and then get yelled at because you haven’t done infrastructure-as-code, but click the button is still possible. I feel like this is on that trendline based upon what you’re saying.

Ram: Absolutely. And the more we can do here, both Pinecone and the broader community, I think the better, the faster the adoption of this sort of technology is going to be.

Corey: I really want to thank you for spending so much time talking me through what it is you folks are working on. If people want to learn more, where’s the best place for them to go to find you?

Ram: Pinecone.io. Our website has a ton of information about Pinecone, as well as a lot of standard documentation. We have a free tier as well where you can play around with small data sets, really get a feel for vector search. It’s completely free. And you can reach me at Ram at Pinecone. I’m always happy to answer any questions. Once again, thanks so much for having me.

Corey: Of course. I will put links to all of that in the show notes. This promoted guest episode is brought to us by our friends at Pinecone. Ram Sriharsha is their VP of Engineering and R&D. And I’m Cloud Economist Corey Quinn. If you’ve enjoyed this podcast, please leave a five-star review on your podcast platform of choice, whereas if you’ve hated this podcast, please leave a five-star review on your podcast platform of choice along with an angry, insulting comment that I will never read because the search on your podcast platform is broken because it’s not using a vector database.

Corey: If your AWS bill keeps rising and your blood pressure is doing the same, then you need The Duckbill Group. We help companies fix their AWS bill by making it smaller and less horrifying. The Duckbill Group works for you, not AWS. We tailor recommendations to your business and we get to the point. Visit duckbillgroup.com to get started.

Announcer: This has been a HumblePod production. Stay humble.