Sponsored by: Fivetran | Building an AI-Ready Data Foundation at Superhuman with Databricks

Summary

Superhuman standardized third-party data ingestion on Fivetran to simplify data movement from Salesforce, Stripe, Zendesk, Outreach, Pardot, and Qualtrics into the Databricks Data and AI platform, enabling go-to-market and finance teams to access trusted data without custom engineering requests.
Using Fivetran's Connector SDK with AI-assisted development, Superhuman reduced specialized connector development from months to days, contributing to faster reporting for a $300 million ad program and a 10% lift in customer lifetime value.
The self-serve operating model built in Databricks allows multiple business functions to access governed third-party data at scale, supported by Databricks Genie agents injected into the pipeline for automated data quality validation.

Chapters

FAQs

Why did Superhuman standardize on Fivetran for data ingestion into Databricks?

Superhuman needed to consolidate data from many third-party SaaS systems — including Salesforce, Outreach, Pardot, Stripe, Zendesk, and Qualtrics — without turning every new data request into weeks of custom engineering. Standardizing on Fivetran provided over 800 pre-built connectors and a Connector SDK that enabled the team to build and maintain integrations far faster than managing proprietary pipelines.

What results did Superhuman achieve with this data foundation on Databricks?

According to this video, Superhuman achieved faster reporting and alerting for a $300 million ad program, enabled internal AI-ready workflows, and contributed to a 10% lift in customer lifetime value. Connector development time was also reduced from months to days by using the Fivetran Connector SDK with AI-assisted development.

How does the Fivetran Connector SDK work for building custom data pipelines?

The Connector SDK allows teams to build API-based connectors by specifying endpoints, rate limits, and sync configuration, with the resulting connector running as a native Fivetran pipeline. This video includes a walkthrough of the connector setup and local testing workflow, demonstrating how to validate syncs before deploying to production on Databricks.

How does Databricks Genie fit into the Fivetran ingestion pipeline?

The architecture shown in this video includes injecting Databricks Genie agents into the Fivetran pipeline for data quality validation, using natural-language queries to verify consistency and accuracy as part of the ingestion workflow. This allows teams to combine automated data movement with AI-powered quality checks in a single governed pipeline.

Full transcript

[00:09] Good afternoon everybody. Sound okay? All right. >> Well, hey, uh, George and I are really happy to be here with you this afternoon. Thank you for joining us. We're going to be talking about >> some really interesting things that Superhum has been doing to build an AI
[00:21] ready data foundation with both Fiverr and data bricks. Before we get into that though, quick introductions. I'm Kelly. He's George. I look after partner sales engineering at Fiverr. I always tell
[00:34] people I am here for one simple reason. I've been here four years. By the way, one simple reason though. Prior to this four years, I recommended Fiverr probably 100 125 times to clients when I
[00:48] was doing data consulting. So, it made it for an easy transition for me. Again, really happy you're here. George, quick intro. >> Yeah. So George Korea, I've been a software engineer at formerly known as
[01:00] Grammarly. Now it's at Superhuman. I've been there for about four years and previously before that had experience at Dropbox. Both kind of cutting edge data engineering teams and now we're just focused in on how do we build really performant thirdparty data platforms
[01:13] with Fiverr. >> Really good. Really good. George, great to be here with you. >> Yeah, great to be here. >> So normally when I'm doing one of these sessions with a large audience, typically I'll have a poll. I don't have a poll today, but it would be
[01:26] interesting to know how many of you are fairly familiar with Fiverr or maybe a better way of saying it, how many are being introduced to Fiverr for the first time today? >> Yeah. Okay. Usually it's about 40 or
[01:39] 50%. So you guys can see the slide. Here's how I think of Fiverr. Been around since 2012. Late stage venturebacked. We have been kind of help coin the term modern data stack. Kind of
[01:53] moving away from that. these days with the with the way that agents are taking over interaction with data, with systems, with really everything that we're interacting with these days. I mean, for me personally, George, I
[02:06] typically have my hands on one keyboard and I've got about five clawed code sessions going at one time. Probably the same. Yeah. >> So, it's really taken over. So, we're really starting to talk now a lot more
[02:18] about D uh agents as the primary data consumers, customers across the board, over 8,000, every industry, every geo, every size.
[02:31] So, you can see some of the names up there. We did get I think it's public now. I don't know. Sydney, may uh our data bricks award good? Yes. So, I think we've got a 2025 up there. We were awarded ISV
[02:44] customer centric partner, something along those lines, customer success partner of the year. So another, we're really pleased. Another uh really cool data bricks award as a datab bricks ISV
[02:56] partner. All right. So for those of you again, a lot of hands in the audience that had not seen or heard of Fiverr before, think about us across four major dimensions. So you want to move data.
[03:09] We've got a catalog of almost 800 sources that moves data into the datab bricks lakehouse. Think databases, SAS applications, file systems, and event systems. We can then manage that within
[03:23] data bricks. We also have a management uh solution that moves it into a lowcost lakehouse as well that you can plug data bicks compute into. We've also merged with anybody know DBT labs? DBT, we've
[03:38] merged with DBT Labs. That is official. Closed a couple of weeks ago. So, think anything you need to go do to go from raw bronze all the way through silver gold. And then we also activate data as
[03:50] well. There's a lot in there we can talk about, but I want to hit a couple of the high points. Lastly, and then I'm going to turn it over to George. We've been working with the data bricks ecosystem for a long, long time. Over 500
[04:02] fantastic customers. hopefully a few in the audience here. Uh today I mentioned the uh the new award that we got and when you think about Fiverr and data bricks together the the word for me that
[04:14] pops out is optionality. If I'm a small medium or certainly an enterprise size customer I need options today. I need to be able to support data bricks data applications uh classic
[04:28] workloads and now agentic workloads with everything going on in the genie space. Fiverr makes that really easy. That was quick. We're happy to talk to you further afterwards, but really want to get over and George, turn it over to you
[04:42] here. Tell us a little bit about where Superhum is and the journey for your company today. >> Yeah. Yeah, sure. So, basically at Superhum, we were previously known as like Grammarly. Uh, so Grammarly is a pretty well-known brand. A lot of users
[04:55] already are aware of like how to use like NLP models to really expand on like how to actually be a more performant writing assistant. with the whole change of like open AI anthropic coming in that really challenged us and so what we
[05:07] ended up doing is we ended up focusing on like acquisitions um to see if we can really expand on our on our bread and butter of actually building AI models for like the common person um so we bought two uh major acquisitions which
[05:19] is Kota and also superhuman male. So Kod superhuman mail and grammally alto together has really built this new rebranding of what we call superhuman. Essentially want to be an assistant to actually enable you to be a superhuman to uh engage with all the third party
[05:33] apps that you possibly have through like kod building API tools uh the grammarly AI assistance those are really popular with like students. So basically having that focus of like the funnel of like getting users to really be aware of our
[05:45] branding and then also being a good partner in your enterprise tool sets. >>
[05:58] >> Yeah. >> Yeah. So it's been fantastic for Copa with Atlan recommended to me. >> Oh yeah. >> Yeah. Great. That's awesome. Yeah. Superhuman mail has been like has a bit of a learning curve but once you really get uh attuned with it it's like really helpful and really really powerful. Um
[06:11] but awesome. So inside our stack this is probably like what most like data engineers are really focused on. um how do we actually like partner up with Fiverr and expand it? So our main story kind of starts with like how how many data sources do we really have? We so
[06:25] I've been like the lead data engineer for our finance teams, sales teams, our marketing teams especially. Um and pulling from these data sources and working with like non-technical stakeholders has been a challenge. Uh so some people think of like you know what
[06:38] I can just build a airflow connector myself. I can just use cloud code just run it myself. Um that has bandwidth that has like a maintenance fee on top of that. And so you really have to be aware of like how many engineers do you really want focused in on just building
[06:50] custom connectors when you can have an AI code or just expand it out to like 5. They've they've had these massive like libraries. I think it's like 700 plus connectors. Is that is that right? >> Yeah, we're we're pushing 800 now. >> Okay. Pushing 800. that number keeps
[07:02] going up and up and and that's a good thing cuz then if you end up uh purchasing like 5 partner up with them um those partnerships end up building over time and you build trust over time and that's really what is powerful to the story is that we have these data
[07:14] sources we have all the latency it works in natively within our AWS uh ecosystem and lands within data bricks is like delta lakes uh we categorize that all with like unity catalog and we have some
[07:26] type of like management over like permissioning on top of unity catalog but it makes that really easy and really streamlined for our users, nontechnical users specifically, to really expand on their data sources and they're able to move quickly on reporting. I'll I'll
[07:39] highlight one major thing. Um let's see if we have this in here. Yeah, one major thing is like these custom pipelines. So there's as you can see there's many vendors at here at the data bricks conference. Um and you can imagine that
[07:52] we have you know 100 plus we're actually about pushing 200 different pipelines >> right now. So we're expanding on that. So that's really enabled us to do like u performance survey analysis like zenes data performant ad spend analysis. Uh we
[08:06] have a probably hundreds of millions of dollars in ad spend a year and so having proper reporting of that is like really essential for us. Um and then the major things that we have really focused in on like how can we use AI with like Fiverr.
[08:19] So we actually had a pretty good meeting with them. Uh shout out to our account manager Chris Grahams there. uh he's been really pushing on like hey how can we be a good partner for like this agentic AI workflows activations is a new space as well um so we kind of
[08:31] zeroed in on like hey we want to be able to build custom pipelines regardless if it's provided by fiverran or if it's open source like we can we use cloud code to user fiverr SDK so >> so if it's not in that 800 connector
[08:44] catalog this is what you're talking about because I saw in your previous slide you had a plus connector SDK >> exactly okay >> yeah exactly so if you if we go if I go back really that plus connector SDK. So, anything that's provided by Fiverr, easy
[08:57] to use. Provide your credentials. But if it's not, the connector SDK plus this cloud skill that we've developed makes it just as easy. You provide a API documentation, API keys, put into your code repo, and you'll be able to locally
[09:09] test it and then push it into Fiverr as if it was a native connector. Um, do we have a demo? Great. Should we pop into the demo? I >> while you're getting set up, I'll switch you over. Sweet.
[09:21] >> You ready? >> Yeah. Good. All right. So, >> hi everybody. So, today we're going to be going over a demo of how to use a cloud skill called create fivet
[09:34] connector to essentially build a fiverr third party connector with fiverr SDK. So, here we have a a create fiverr connector skill. You can see this skill right here outlined being all the steps that we need to do to actually build this connector. One is basically provide
[09:47] the central links which is like API documentation. Essentially you want to like document the how to off rate limiting. Next thing is also the data data shape as in like which endpoints you want to call. How do you actually
[09:59] pull in different data? Is it pagionation? Is it incremental configuration over the connectors as in do you need to pass like API keys sensitive information? What are the primary keys etc. So you can see all these like steps are being outlined and
[10:12] saved so you don't have to like keep redoing things that essentially are going to be done on the engineering teams. So this saves you a lot of time. So an example of how we actually use this. Let's let's use this to build a uh third party connector for our survey
[10:24] data called sprig. So I called the skill. It's going to ask me for documentation. Provide a link there. So docs.sprig with all the references. Great. And now it has an outline of like what we actually need to provide. So in
[10:36] this case, I'm going to provide an API key, the base URL, and also the rate limiting, which is 1,000 requests per second. It does a lookup of all the endpoints that are available to me. And it's also going to ask me like which which responses or which endpoints I want to pull from. Um so in this case I
[10:49] only want to pull from sync responses and survey data and we're going to go with an incremental approach. So it's going to read that does all the lookups. After all the lookups it sees that there's
[11:01] like filtering that we can provide on milliseconds. Each API is slightly different. So at least the cloud skill is going to have that inference and basically double check with you if it's going to be something that's like appropriate. Some cases there might be
[11:13] some extra handling that you can do like custom handling which you can also provide. In this case I only want to pull in those two data points which is response and survey. I'm going to pass the primary key as a response group U ID and ids for survey.
[11:27] Okay. So now that we have the primary keys and the objects identified. Now it's going to actually implement the scaffolding which is happening all here. Scroll through that. Um so now this the connector has a scaffolding. It's now located in our registry. Now we
[11:40] essentially need to test this. So we have some API key some sensitive information you go to the configuration JSON you can pass this I've exed out my API key in your case you can provide that there for you and then I basically say okay it has local testing you can do
[11:53] it manually or I just tell cloud itself you can do uh local testing install has a Python library
[12:05] couple things that need needs to be adjusted um as we go all the way through it's adjusting it's checking in with you and then lastly we have a successful sync. This is all being done locally. Once this is committed then we actually have a
[12:17] continuous deployment process. But from here on out this is with local testing and with the watermarking this essentially proves out that we have a full endto-end connector running with using five SDK all being created by
[12:30] cloud. >> Cool. Yeah. Hopefully that was uh informative. I know looking at cloud sessions, they tend to be uh
[12:48] so there there are a lot of ways to build a custom connector. Quick quick highlights. Didn't you guys you showed one way there? >> Uh there's a massive connector example library out there as well. Do you use
[13:00] some of those as kind of templates? >> Yeah. Yeah. So templates. So, one thing to look out, so if you look up Fiverr connector SDK, there's going to be a a GitHub repo that actually has all these connections. So, MD files are actually shared there. That was a more custom
[13:13] tweaked uh MD file specific to to superhuman. Um, but there's actually a community. So, people are actually posting out different connectors. So, you can use one of those as like scaffolding uh for existing ones. So, think of things like, hey, there's a connector with pagonation or a connector
[13:26] with incremental syncs or a another connector that maybe is a bit more custom. Um you can there's all these different templates so you can easily just like grab one of those, grab a link and just post it into the cloud uh cloud
[13:38] skill and it should just be able to pick up that context. >> Yeah, one of the ones that I contributed out there uh I injected two Genie agents into the pipeline >> to really take a look at the data source before I ever moved anything because a
[13:51] lot of times you think about well this custom connector is just going to take data and move it. In this case, I had two Genie agents debate about what would be for this use case the appropriate uh
[14:03] schemas and tables and different columns that we would need to achieve the result and then I had a moderator agent that made the final decision all all done with Genie and then move and then moved it into data bricks from there. So
[14:17] >> it's kind of you have an open pallet with connector SDK. >> Yeah. Yeah. Here's one example that we actually did is that there's an existing partardot. So uh product connector which is a Salesforce product um it was high
[14:29] volume there was a lot of filtering it could have been it could not have been done with the native connector so we ended up taking advantage of the c the connector SDK and adding in custom filters so we actually reduce the volume down so it actually became a more smooth line uh third party pipeline as opposed
[14:42] to like bringing everything in having to filter having to do almost like a medallion model doing cleanup cleansing filtering it's a lot of overhead with five SDK it made a lot easier so >> and George when I think about custom connectors too. I typically think about
[14:55] it in maybe three big chunks. You've got to interact with the source, create whatever that thing is that you want to do with the source. I've got to process the data in the middle, data, type it out, those kind of things. And then I
[15:07] need to move it somewhere. I need to move it in the right way into data bricks or into a data lake. >> By train connector SDK, you really just have to do that first piece because when I when I saw you, you know, say, hey,
[15:19] it's ready to deploy. Once you deploy those other two big chunks, they're already done. >> Yeah, exactly. Yeah, it's all being done on the Fiverr cloud. So basically, if you need to do like schema management, schema like merges, it's all being done
[15:32] five. So that makes it really easy. >> Yeah, very favorite thing about it. >> Well, let me ask you this. You guys have had a ton of things going on. You've got company shifts, everybody coming together. You got new technologies,
[15:45] you've got this >> AI uh agentic age that is uh really advancing quickly. What do you have come? I'm just going to go between now and the end of the year. Forget next year. Let's say next six months. >> Yeah. Yeah. I think what we want to do
[15:58] is essentially can we put these tools into the hands of our non-technical stakeholders as in we still get tickets for like hey can you build this like third party connector? Sure, we have the cloud skill, but think of things like tools like uh like linear almost like a
[16:10] jerk ticket. A marketer puts in, hey, I want to build from this third party source, have an agent look that up. We have modern agents, data quality agents, being able to interact with them in a like almost like a real-time queue. You
[16:22] take in all these uh jer tickets and then the engineer just works as like reviewing the code. That's like the ideal state. And so we want we have pieces of that already. Um, basically going through like agentic data
[16:34] operations like data quality testing. That's like one thing we really want to like partner up with. Um, and then also you mentioned like activations, right? And so like activations which is essentially how do you push in that data that you have in data bricks into
[16:46] thirdparty tools. Um, that's something that we are looking into like can we actually build like aentic workflows to not only just pull in data but also push out data. >> Yeah. And take action. Now I've I have the same kind of thing. I mean today Fiverr supports a couple of hundred of
[16:59] those destinations on the activation side. And then you think about as you were talking, >> what can I do from an agent standpoint to not just move but actually take some action as things are happening as well.
[17:12] So potentially moving to an agent or to a system or some combination. >> Yeah. Yeah. Exactly. Yeah. Yeah, I think data discovery is going to be another tool like we've actually worked with uh with like Fiverr to build a more like
[17:24] data governance model as then we have like data uh like restricted or more sensitive data sets also like on once those data sets land in data bricks finding those data sets what's available sometimes we have to tweak it and so I think like in the future we're going to
[17:37] actually have a bit more an intelligent system we have an agent to actually say hey we just have this new third party source pulling in maybe it's coming in for marketing but it could be useful for finance so having a way to actually announce that in an automated in a fashion. Yeah.
[17:49] >> All right. Well, as we're wrapping up, anything I may I may be throwing this off the cuff to you, but any recommendations for everybody? You say, "Hey, I'm I'm starting to get rolling in this data bricks world here. Maybe it's,
[18:01] you know, something else." Any recommendations, your learnings over the past year or so that you could say, "Hey, keep this in mind or don't do this or you really do want to do this?" >> Yeah, that's a really good question. Um,
[18:20] I think the number one thing that people almost like trip up on is how to test the data quality of like third party sources. Um, because a lot of nontechnical people end up pulling in data as much as possible and they might think like, oh, this data is actually
[18:32] resulting in an error or the schema is not exactly what I want it to do. And people tend to think, oh, it's it's a fetch problem. Uh sometimes it's actually more of a thirdparty API problem as in like maybe there's rate limits or maybe you're not familiar with
[18:45] like the data model and so there's going to be some overhead into like how do you actually partner up with these third party sources um because I think a lot of folks like end up losing trust >> in in 5chan just because they don't have
[18:57] information of like how to actually interact with the API itself using cloud to interact hey tell me exactly how this API can pull in data what exactly does it have that's been really powerful cloud has almost been like that instead
[19:09] of interacting with like a customer support agent from a third party vendor, I can just tell Claude, hey, have the API docs, explain it to me, be your expert on this, and then it's a lot smoother because then I'm interacting with an agent that has all the same
[19:21] documentation and if there is an issue, then I go to the customer support, but at least I'm not wasting time with like very basic documentation or like maybe it's like a little golden nugget hidden in some p random page. >> Love it.
[19:33] >> Well, we really really appreciate it, George. Thank you so much. Uh thank you to everyone who attended today. Uh feel free, we're just across the data bricks uh booth there. Uh 420 and 4:30 Fiverr
[19:45] and DBT labs together. Come talk to us. George and I can hang out for a while here as well. Thank you again. Really appreciate it. >> Thanks everybody.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.